# Speech Recognition

> Source: https://aiwiki.ai/wiki/speech_recognition
> Updated: 2026-08-01
> Fact-checked: 2026-07-29
> Categories: Deep Learning, Machine Learning, Natural Language Processing, Speech & Audio AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Speech Recognition." aiwiki.ai, 1 Aug 2026. https://aiwiki.ai/wiki/speech_recognition
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Speech recognition, usually called automatic speech recognition (ASR), is the computational task of converting a spoken-language signal into a sequence of written symbols. The input may be a live microphone stream or a recorded file. The output may be plain text, a time-aligned transcript, or a stream of partial hypotheses with confidence information. ASR is a problem in [Machine Learning](https://aiwiki.ai/wiki/machine_learning), speech processing, and [Natural Language Processing](https://aiwiki.ai/wiki/natural_language_processing), but it is not the same task as identifying a speaker, separating speakers, translating speech into another language, or inferring a user's intent.

A recognizer must account for several kinds of variation at once. Different speakers can express the same words with different pronunciations, rates, pitches, and disfluencies. Microphones, rooms, codecs, noise, reverberation, and overlapping speech alter the recorded signal. Languages also differ in sound inventories, writing systems, word boundaries, and conventions for numbers, names, and punctuation. Modern systems learn much of this mapping from data, while deployed pipelines often add separate components for segmentation, decoding, text normalization, confidence estimation, and speaker attribution.

No single accuracy figure describes speech recognition in general. Results depend on the language, domain, acoustic conditions, transcript conventions, test-set composition, and scoring procedure. A low error rate on read audiobooks does not establish equivalent performance for a noisy meeting, accented conversation, clinical dictation, or code-switched speech. For the same reason, public benchmark results should be read as results for a specified system and evaluation set, not as a universal measure of human or machine transcription ability. A separate article surveys [Automatic Speech Recognition Models](https://aiwiki.ai/wiki/automatic_speech_recognition_models); this article focuses on the task, its historical development, main methods, evaluation, and deployment constraints.

The way recognizers are built and packaged has changed considerably since the early 2020s. As of July 2026 three overlapping families are in common use: dedicated audio-in, text-out encoder-decoder models; hybrid systems that attach a speech encoder to a pretrained text [Large Language Model](https://aiwiki.ai/wiki/large_language_model); and general-purpose [Multimodal](https://aiwiki.ai/wiki/multimodal_ai) models for which transcription is one capability among many. The third family has changed what a "speech recognition product" means, because a conversational model can consume speech and act on it without exposing a transcript at all. The methods below are described first, followed by the current state of the field, which moves faster than the methods do.

## Task and system boundary

The core ASR problem can be stated as finding a likely token sequence `W` for an observed acoustic sequence `X`. A statistical recognizer commonly expresses the decision as:

```text
W* = argmax_W P(X | W) P(W)
```

Here, `P(X | W)` represents how well a candidate word sequence accounts for the acoustic observation, while `P(W)` represents a prior preference over word sequences. In a modular system these terms are associated with an acoustic model and a [Language Model](https://aiwiki.ai/wiki/language_model). End-to-end neural systems may not expose the same components, but decoding still balances evidence from the audio with learned regularities in output sequences.

The target transcript must be defined before training or evaluation. It may preserve hesitations, partial words, capitalization, punctuation, and non-speech events, or it may normalize them away. A system trained to produce lowercase lexical words is solving a different output task from a system expected to emit display-ready prose with punctuation and inverse text normalization. The transcript policy determines what counts as an error.

Several adjacent tasks are often combined with ASR but remain conceptually separate:

- **Speech activity detection** decides which time spans contain speech. A survey of implementations appears in [Voice Activity Detection Models](https://aiwiki.ai/wiki/voice_activity_detection_models).
- **Speaker diarization** assigns speech spans to anonymous speaker labels, answering "who spoke when."
- **Speaker recognition** associates a voice with an enrolled or hypothesized identity.
- **Speech translation** maps speech in one language to text in another.
- **Spoken language understanding** predicts intents, entities, dialogue acts, or other meanings from speech.
- **Forced alignment** aligns a known transcript with an audio recording rather than discovering the transcript from scratch.

An application can cascade these tasks or train a joint model. A meeting-transcription service, for example, may detect speech, separate or diarize speakers, recognize words, restore punctuation, and format numbers. Reporting only the recognizer's word error rate does not measure the quality of every stage.

## Historical development

Early speech recognizers handled small vocabularies under tightly controlled conditions. In 1952, Davis, Biddulph, and Balashek described a system for recognizing spoken digits over telephone-quality input from one individual. The paper reported high accuracy only after adjustment to that speaker and task, so it should not be read as evidence of general, speaker-independent recognition.[1] IBM's Shoebox demonstration, introduced publicly in 1961 and shown at the 1962 Seattle World's Fair, recognized the ten digits and six command words. It coupled speech input to a calculator, illustrating command-and-control recognition rather than open-ended transcription.[2]

Research in the 1970s increasingly treated recognition as statistical sequence decoding. Jelinek's 1976 account described a system organized around acoustic measurements, probabilistic models, linguistic constraints, and search.[3] [Hidden Markov Models](https://aiwiki.ai/wiki/hidden_markov_model) became a central framework because they represented a word or subword as a sequence of latent states and supported efficient dynamic-programming algorithms. Rabiner's 1989 tutorial consolidated the forward-backward, Viterbi, and parameter-estimation methods used in HMM-based recognition.[4]

Large-vocabulary, continuous, speaker-independent recognition became a major research goal during the 1980s. Carnegie Mellon University's Sphinx work combined statistical acoustic modeling, lexical pronunciation information, language modeling, and search; the university's project history places the first Sphinx system in 1988 and identifies Kai-Fu Lee, Raj Reddy, and Roberto Bisiani as central contributors.[5] This was a milestone in a continuing research program, not the beginning of all speech recognition.

By the 1990s and 2000s, modular recognizers commonly combined context-dependent phone HMMs with [Gaussian Mixture Models](https://aiwiki.ai/wiki/gaussian_mixture_model) for acoustic likelihoods, pronunciation dictionaries, and [N-gram](https://aiwiki.ai/wiki/n-gram) language models. Weighted finite-state transducers provided a common representation for acoustic context, lexicons, grammars, and alternative hypotheses. Composition, determinization, minimization, and weight pushing made it possible to assemble and optimize decoding graphs systematically.[6]

Research toolkits made complete recipes more reproducible. The Hidden Markov Model Toolkit, or HTK, distributed source code and documentation under a registration and license agreement; its official site should not be used to imply that it has the same open-source terms as Apache-licensed software.[7] Kaldi, described in 2011, provided an Apache 2.0-licensed C++ toolkit using finite-state transducers, acoustic modeling libraries, and experiment recipes.[8] Toolkits are implementations, not fixed recognition models, and their results depend on the data and recipes used with them.

Neural networks first entered many production-quality systems as replacements for the Gaussian-mixture observation model inside an HMM decoder. Hinton and colleagues' 2012 review reported that deep feed-forward networks used as acoustic posterior estimators outperformed strong Gaussian-mixture baselines across several speech-recognition groups and tasks.[9] These systems remained hybrid DNN-HMM systems: the [Deep Neural Network](https://aiwiki.ai/wiki/deep_neural_network) supplied acoustic scores, while HMM state structure, a lexicon, a language model, and a decoder remained explicit.

End-to-end approaches reduced the number of separately trained modules. Connectionist temporal classification enabled direct sequence labeling without frame-level alignments.[10] Recurrent CTC systems were then evaluated on large-vocabulary speech recognition without an intermediate phonetic target.[11] [Attention](https://aiwiki.ai/wiki/attention)-based encoder-decoder models learned a conditional mapping from acoustic frames to output characters,[12] while recurrent neural network transducers modeled a distribution over monotonic alignments between input frames and output tokens.[13] Later architectures combined convolution, recurrence, and self-attention. The Conformer, published in 2020, joined convolutional modules for local patterns with self-attention for broader context and reported paper-specific improvements on LibriSpeech.[14]

Another development was large-scale pretraining. [wav2vec 2.0](https://aiwiki.ai/wiki/wav2vec) learned representations from unlabeled speech by masking latent inputs and using a contrastive objective, then adapted the model with labeled transcripts.[15] HuBERT instead used clustered hidden-unit targets for masked prediction.[16] Whisper trained an encoder-decoder model on 680,000 hours of weakly supervised multilingual and multitask data collected from the internet, then evaluated transfer without dataset-specific fine-tuning on multiple benchmarks.[17] These projects illustrate different training regimes. They should not be collapsed into one method or treated as proof that labeled, domain-specific evaluation is unnecessary.

Two further shifts followed. From roughly 2023, several groups began attaching a speech encoder to an already-trained text language model through a small adapter, so that the component producing words is a general-purpose text model rather than one trained only for transcription. By 2026 systems of this shape occupied much of the top of the main public accuracy leaderboard.[31] Separately, full-duplex speech-to-speech models appeared that generate audio directly from audio and use an interleaved text stream only as an internal scaffold, so no transcript is necessarily produced for the user.[40] Neither shift retired the earlier designs. CTC and transducer decoders remain the fastest option for bulk transcription, and the same benchmark suites are used to compare every family.[31]

## Audio representation and training data

A waveform is a sequence of sampled amplitudes. Recognition systems usually operate on shorter, overlapping regions because speech changes over time but is approximately stable over a small interval. Traditional front ends calculate spectral representations such as log Mel filterbank energies or Mel-frequency cepstral coefficients. The resulting feature sequence compresses some waveform detail while preserving information useful for distinguishing speech sounds. Modern encoders may still consume filterbanks, or they may learn a front end directly from waveform samples.

Feature normalization and augmentation address nuisance variation but do not eliminate it. A pipeline may normalize mean and variance, perturb speed, mix noise, simulate reverberation, or mask regions of a spectrogram. SpecAugment applies time and frequency masks to filterbank features and was evaluated as a simple regularizer for end-to-end ASR.[18] Its results are evidence for the paper's tested models and datasets, not a guarantee that every masking policy improves every recognizer. [Data Augmentation](https://aiwiki.ai/wiki/data_augmentation) must preserve the transcript and resemble conditions the system may encounter.

Training examples pair audio with a reference transcription. Useful metadata may include language, locale, speaker, recording device, environment, license, consent status, and transcript provenance. Segment boundaries matter: excessively long or incorrectly cut examples make alignment harder, while boundaries that consistently remove hesitations or silence may produce a mismatch with live audio. Duplicate speakers or recordings across training and test partitions can also overstate generalization.

Public corpora differ substantially:

| Corpus | Speech and scope | Appropriate interpretation |
| --- | --- | --- |
| LibriSpeech | Approximately 1,000 hours of 16 kHz read English derived from LibriVox audiobooks | A well-defined read-speech benchmark, not a proxy for every English domain[19] |
| Common Voice | Crowdsourced, community-validated recordings across many languages and speaker backgrounds | A multilingual collection whose releases and per-language composition change over time[20] |
| FLEURS | Parallel read speech in 102 languages, built from the FLoRes-101 text benchmark, with roughly 12 hours per language | A multilingual evaluation resource for ASR and related speech tasks, not a large training set for every language[21] |

[LibriSpeech](https://aiwiki.ai/wiki/librispeech) is especially common in architecture papers because its splits and scoring setup are familiar. That convenience creates a risk of benchmark overgeneralization. Read audiobooks have different speaking styles, acoustics, and lexical distributions from meetings, call centers, children's speech, medical conversations, or field recordings. A deployed system needs test data sampled from its actual operating conditions.

The mapping from transcripts to model targets is also a modeling choice. Traditional systems often use context-dependent phones and a pronunciation lexicon. Neural systems may predict characters, bytes, graphemes, phonemes, words, or subword units. [Tokenization](https://aiwiki.ai/wiki/tokenization) affects vocabulary size, sequence length, treatment of unknown words, and behavior across scripts. A character inventory can avoid an out-of-vocabulary word list but may create long output sequences. Subwords shorten sequences and share parts across words, but their segmentation reflects the training text. Phoneme targets can share pronunciations across spellings but require a pronunciation mapping and a later path to written words.

Transcript normalization must be specified for both training and scoring. Decisions include whether "twenty-one" and "21" are equivalent, how abbreviations are expanded, whether filled pauses count, and how scripts without whitespace are segmented. If references and hypotheses use different normalization rules, the measured error can reflect formatting rather than acoustic recognition.

## Modular statistical systems

A modular recognizer separates knowledge sources so they can be trained, inspected, and replaced independently. Its major components are:

1. **The acoustic front end** converts the waveform into frame-level features.
2. **The acoustic model** scores speech units or HMM states for those frames.
3. **The pronunciation lexicon** maps words to one or more phone sequences.
4. **The language model** scores candidate word sequences.
5. **The decoder** searches for high-scoring paths and can return a best transcript, an N-best list, or a lattice.

In a classic HMM-GMM system, the HMM expresses state transitions over time and Gaussian mixtures describe feature distributions within states. Context-dependent units account for the fact that a phone's realization changes with neighboring phones. Speaker or environment adaptation can transform features or model parameters using enrollment or first-pass hypotheses. A hybrid DNN-HMM system retains the state sequence and decoding graph but uses a neural network to estimate state-related scores.[4][9]

The lexicon makes pronunciation assumptions explicit. It can list multiple pronunciations for a word and assign probabilities or weights. It can also generate pronunciations for new words with a grapheme-to-phoneme model. This separation is valuable in domains with names, acronyms, or specialist terms, but errors can arise when the lexicon omits a valid pronunciation or the acoustic model has little evidence for it.

The language model contributes information that is not reliably recoverable from acoustics alone. Homophones and near-homophones can have similar acoustic evidence but very different sequence probabilities. In a modular decoder, an acoustic scale and insertion penalty control how acoustic and linguistic scores interact. These values are tuned on development data. Increasing the language-model weight can improve plausible completions while also making the decoder more likely to override unusual but correctly spoken words.

Weighted finite-state decoding represents the pieces as compatible transducers and composes them into a searchable graph. This design can keep alternative paths in a lattice for later rescoring with a stronger language model.[6] A lattice is useful when downstream processing needs uncertainty rather than one irreversible string.

Modularity has practical advantages. Text data can update a language model without retraining the acoustic model, and a pronunciation can be added without collecting new speech. Components and failure sources are comparatively visible. The tradeoff is engineering complexity: separately estimated pieces optimize different objectives, and independence assumptions may not match real speech. End-to-end systems were developed partly to learn more of these interactions jointly.

## End-to-end objectives and architectures

"End-to-end" does not denote one architecture. It usually means that a neural model is optimized to map acoustic input to transcript tokens without requiring frame-level phone labels or a separately trained HMM-GMM acoustic model. Some end-to-end systems still use an external language model, a lexicon, a text normalizer, or a second-pass decoder.

### Connectionist temporal classification

[Connectionist Temporal Classification](https://aiwiki.ai/wiki/connectionist_temporal_classification), or CTC, defines a probability over output sequences by summing over frame-level paths. Its label alphabet includes a blank symbol. After a path is generated, repeated labels are collapsed and blanks are removed. Dynamic programming marginalizes all paths that collapse to the reference sequence, so training does not require a known alignment between each label and input frame.[10]

CTC makes a conditional-independence assumption between output labels given the encoded input. The encoder can still use broad acoustic context, but the output layer does not directly condition each label on previously emitted labels. Decoding may be greedy, or it may use [Beam Search](https://aiwiki.ai/wiki/beam_search) with a lexicon or external language model. CTC is also used as an auxiliary loss in systems whose main decoder uses another objective.

### Attention-based encoder-decoder

An attention-based encoder-decoder transforms audio into hidden states and generates output tokens one at a time. At each step, the decoder uses previous output tokens and an attention-weighted summary of encoder states. Listen, Attend and Spell used a pyramidal recurrent encoder to shorten the acoustic sequence and an attention-based recurrent decoder to emit characters.[12] Unlike CTC, the decoder directly models dependencies among output tokens.

Standard full-context attention can use the entire utterance and does not inherently require a monotonic left-to-right alignment. This is useful for offline transcription but complicates bounded-latency streaming. Long recordings can also exceed the context or memory assumptions used in training, so practical systems segment audio or use architectures designed for long inputs.

### Recurrent neural network transducer

The recurrent neural network transducer, or RNN-T, combines an acoustic encoder, a prediction network conditioned on preceding output tokens, and a joint network. Its loss sums over monotonic alignments in a two-dimensional lattice of input time and output position.[13] Because the encoder can be causal and the alignment advances through time, RNN-T is widely studied for streaming. Experiments reported by Rao, Sak, and Prabhavalkar show a streaming end-to-end RNN-T that emits graphemes or wordpieces and can incorporate additional text or pronunciation data.[22]

RNN-T is not automatically low latency. Encoder look-ahead, feature framing, endpoint detection, beam size, hardware, and network transport all contribute. Its prediction network also learns transcript regularities, so domain terms may still need contextual biasing or adaptation.

A later variant, the token-and-duration transducer, predicts a token together with the number of frames to skip before the next prediction, which reduces the number of decoder steps per second of audio. NVIDIA's [Parakeet](https://aiwiki.ai/wiki/parakeet) models use this decoder and report some of the highest throughput figures on public leaderboards.[37]

### Convolution, recurrence, and self-attention

[Recurrent Neural Networks](https://aiwiki.ai/wiki/recurrent_neural_network) process a sequence with recurrent state and can be made causal for streaming. Convolutional encoders efficiently capture local patterns and can downsample long frame sequences. Self-attention relates positions by content and provides broad context, but unrestricted attention over all frames is non-causal and its memory cost grows rapidly with sequence length.

A Conformer block orders a feed-forward module, multi-head self-attention, a convolution module, and a second feed-forward module, followed by layer normalization. The authors' LibriSpeech experiments were designed to capture both local acoustic structure and global interactions.[14] Streaming variants restrict right context, process chunks, cache previous states, or summarize memory. These restrictions create an accuracy, computation, and delay tradeoff that must be measured for the intended device and traffic pattern.

The main objective families have different default properties:

| Family | Alignment treatment | Output dependence | Streaming implications |
| --- | --- | --- | --- |
| CTC | Sums over monotonic blank-augmented paths | Output labels are conditionally independent at the CTC layer | Compatible with causal encoders; external decoding often supplies stronger sequence constraints |
| Attention encoder-decoder | Learns attention between output steps and encoder states | Decoder conditions on earlier outputs | Full-context attention is naturally offline; constrained or chunked variants are needed for bounded latency |
| RNN-T | Sums over monotonic input-output alignments | Prediction network conditions on earlier nonblank outputs | Designed to work with causal or limited-context encoders |

This table is descriptive, not a ranking. Dataset scale, encoder design, decoding, compute budget, and target latency can matter more than the objective label.

## Pretraining and large-scale supervision

Labeled speech is expensive because transcription requires time, language expertise, and a consistent annotation policy. [Self-Supervised Learning](https://aiwiki.ai/wiki/self-supervised_learning) uses an objective derived from the audio itself to learn a representation before labeled adaptation.

wav2vec 2.0 applies a convolutional feature encoder to raw audio, masks spans of latent representations, and trains a Transformer context network with a contrastive task over quantized targets. The authors then [fine-tuned](https://aiwiki.ai/wiki/fine_tuning) the pretrained model with CTC on labeled speech. Their experiments showed large gains in low-label LibriSpeech conditions, but the result depended on substantial unlabeled pretraining data and benchmark-specific decoding choices.[15]

[HuBERT](https://aiwiki.ai/wiki/hubert) creates discrete targets by clustering speech features offline and trains a masked-prediction model to recover cluster assignments in masked regions. The clusters need not be perfect phonetic labels; the learning signal comes from predicting consistent hidden units from context. The published work compared multiple labeled-data regimes on LibriSpeech and Libri-Light.[16]

Self-supervised pretraining and supervised recognition remain distinct stages. A pretrained encoder has not necessarily learned the target writing system, vocabulary, or transcript conventions. Adaptation may attach a CTC head, initialize an encoder-decoder, or provide features to another model. Cross-lingual [Transfer Learning](https://aiwiki.ai/wiki/transfer_learning) can help when languages share acoustic or phonetic structure, but transfer can also favor well-represented languages and domains.

Weak supervision follows a different route. [Whisper](https://aiwiki.ai/wiki/whisper) was trained directly on large-scale paired audio and text collected from the internet, including multilingual transcription and speech translation data. The paper reported 680,000 training hours and evaluated zero-shot transfer across existing datasets.[17] The training labels were not all manually curated to a single standard, so scale trades against label consistency. The model also combines acoustic transcription with learned sequence generation, which is relevant to its robustness and failure modes.

Multilingual scaling can share representations across languages and reduce the need for a wholly separate encoder per language. The [Massively Multilingual Speech](https://aiwiki.ai/wiki/massively_multilingual_speech) project pretrained wav2vec 2.0 models across 1,406 languages and built one ASR model covering 1,107 languages, using readings of publicly available religious texts and language-specific adapters or heads. The JMLR paper reports those language counts and evaluates multilingual transfer, while also documenting the narrow domain of much of its labeled data.[23] Coverage count alone does not establish equal quality across languages.

Meta extended that line of work with Omnilingual ASR, announced in November 2025. The published paper describes recognition for more than 1,600 languages, including over 500 that the authors state had no prior ASR support, with a self-supervised encoder scaled to 7 billion parameters and an encoder-decoder design intended to let communities add an unserved language from a small number of paired examples rather than a full training corpus.[38] The project repository lists four model families released under Apache 2.0 at sizes from 300 million to 7 billion parameters, reports character error rates below 10 for 78 percent of the covered languages, and publishes a companion corpus collected with local organizations.[39] The reported CER threshold is a useful headline but not a usability guarantee: character error rate is measured against the orthography the project chose, and a language with an unsettled writing system can score well on characters while producing text its speakers would not accept.

Multilingual systems face choices about language identification, shared token inventories, script normalization, and negative transfer. A single model can require a language token, infer a language from audio, or produce more than one script. Code-switching is harder than merely supporting each language separately because switches can occur between or within utterances, and mixed-language acoustic and textual training data are scarce. Evaluation should therefore include naturally code-switched material rather than only concatenated monolingual clips.

## Speech encoders, language-model decoders, and omni models

Between 2023 and 2026 the highest-accuracy transcription systems converged on a common shape: a strong acoustic encoder, a small learned adapter, and a decoder that is either a compact task-specific transformer or a pretrained text language model. In parallel, general-purpose conversational models learned to accept audio directly, which moved transcription from a product in its own right to a capability inside a larger system. Both changes are visible in current model cards and leaderboards, and neither eliminates the accuracy, latency, and cost tradeoffs described in the rest of this article.

### Attaching a speech encoder to a text language model

A speech-augmented language model routes encoder output through a projector or adapter into the embedding space of an existing text model, then fine-tunes some or all of the combination. The appeal is that the text model already knows punctuation, capitalization, formatting conventions, named entities, and enough world knowledge to prefer a plausible reading of an ambiguous stretch of audio, none of which has to be learned again from speech data.

NVIDIA's [Canary](https://aiwiki.ai/wiki/canary)-Qwen-2.5B, published in July 2025, is a documented example. It connects a FastConformer encoder taken from canary-1b-flash to a Qwen3-1.7B decoder through a linear projection and low-rank adaptation, totals about 2.5 billion parameters, and is released under CC-BY-4.0 for English. Its card reports an average word error rate of 5.63 and an inverse real-time factor of 418.28 on the Open ASR Leaderboard's English track, and describes two operating modes: an ASR mode that transcribes with punctuation and capitalization, and an LLM mode that recovers the underlying model's text abilities for post-processing but takes text rather than audio as input.[32]

IBM's Granite Speech 4.1 2B, released on 29 April 2026 under Apache 2.0, follows the same pattern with different parts: a 16-layer conformer encoder carrying dual CTC heads for characters and subword units, a two-layer window query transformer projector, and a fine-tuned Granite 4.0 1B text model. Its card reports a mean WER of 5.33 and an RTFx of 231.29 on the same leaderboard for English, French, German, Spanish, Portuguese, and Japanese, and lists keyword-list biasing, punctuation, and truecasing as trained capabilities.[33] Alibaba's [Qwen3](https://aiwiki.ai/wiki/qwen3)-ASR series, released on 29 January 2026 under Apache 2.0, builds on the [Qwen3-Omni](https://aiwiki.ai/wiki/qwen3_omni) stack and ships 1.7B and 0.6B recognizers plus a separate 0.6B forced aligner, covering 52 languages and dialects with language identification and support for singing and music-backed speech.[36]

The design is not universal, and one of the strongest 2026 entries deliberately avoids it. Cohere's Transcribe model, released in March 2026 under Apache 2.0, is described by its developers as a dedicated audio-in, text-out system with a Fast-Conformer encoder holding more than 90 percent of its 2 billion parameters and a deliberately lightweight transformer decoder, precisely so that autoregressive decoding costs as little as possible. Its card reports an average WER of 5.42 across the leaderboard's eight English test sets and 14 supported languages.[34][35] The design choice has consequences that the card states plainly: the model has no automatic language detection, produces no timestamps or speaker labels, and "is eager to transcribe, even non-speech sounds," so it needs preprocessing in noisy conditions.[34]

The Open ASR Leaderboard paper describes the same tradeoff at the level of the whole field. Comparing 86 open and proprietary systems across 12 datasets, the authors report that conformer encoders paired with transformer decoders achieve the best average WER, while CTC and token-and-duration transducer decoders achieve far better inverse real-time factors and are consequently better suited to long-form and batched work.[31] Accuracy and throughput are not the same axis, and a system chosen only on the leaderboard's WER column may cost an order of magnitude more compute per hour of audio.

### Omni models that transcribe as one capability

A second group of systems accepts audio into a general-purpose model. OpenAI introduced gpt-4o-transcribe and gpt-4o-mini-transcribe as dedicated speech-to-text endpoints on 20 March 2025,[43] then made a speech-to-speech model, gpt-realtime, generally available on 28 August 2025 with support for WebRTC, WebSocket, and SIP telephony connections, remote tool servers, and image input.[44] A December 2025 refresh of the smaller audio snapshots claimed, on OpenAI's own noise testing, roughly 90 percent fewer hallucinations than Whisper v2 and roughly 70 percent fewer than the previous gpt-4o-transcribe models, along with lower word error rates on Common Voice and FLEURS without language hints; these are the vendor's figures and no independent replication is cited.[45] The AI Wiki covers these products at [GPT-Transcribe](https://aiwiki.ai/wiki/gpt_transcribe), [GPT-Realtime](https://aiwiki.ai/wiki/gpt_realtime), the [OpenAI Realtime API](https://aiwiki.ai/wiki/openai_realtime_api), and [GPT-Live](https://aiwiki.ai/wiki/gpt_live).

Google's [Gemini Live](https://aiwiki.ai/wiki/gemini_live) API processes continuous audio, video, or text streams in a single model rather than chaining separate recognition, reasoning, and synthesis stages, and exposes text transcripts of both the user's input and the model's output as an optional side channel alongside barge-in interruption handling.[46] Alibaba's Qwen3-Omni uses a mixture-of-experts thinker-talker design that consumes text, images, audio, and video and emits both text and speech.[47] Other entries in this family with pages on the wiki include [Qwen2-Audio](https://aiwiki.ai/wiki/qwen2_audio), Meta's [SpiRit-LM](https://aiwiki.ai/wiki/spirit_lm) and [SeamlessM4T](https://aiwiki.ai/wiki/seamless_m4t), and [GLM-4-Voice](https://aiwiki.ai/wiki/glm_4_voice).

For the task boundary described earlier, this matters more than the accuracy numbers do. When a conversational model answers a spoken question, the transcript may be an internal representation that is never shown, never stored, and never scored. Word error rate then measures a component that no user sees, while the outcome the user cares about is whether the model did the right thing. Evaluation has to follow the product: a voice agent should be measured on task completion, tool-call correctness, and turn-level behavior, with transcription metrics used as diagnostics rather than as the headline. OpenAI's own December 2025 notes reflect this, reporting instruction-following and tool-calling accuracy improvements for the speech-to-speech models rather than WER.[45]

### Open-weight recognizers as of July 2026

The figures below are each developer's published results on the English track of the Open ASR Leaderboard, which averages WER over eight test sets including LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech, TED-LIUM, and VoxPopuli. They are comparable to each other only in that respect. They say nothing about the languages, domains, or acoustic conditions a given deployment cares about, and the leaderboard is live, so the ordering changes.

| Model | Released | Size and design | Languages | License | Published English-track average WER |
| --- | --- | --- | --- | --- | --- |
| [Whisper](https://aiwiki.ai/wiki/whisper) large-v3 | 2023 | 1.55B encoder-decoder, weak supervision | 99 | MIT | Reported by Cohere as 7.44 under the same protocol[35] |
| Whisper large-v3-turbo | Oct 2024 | 809M; decoder reduced from 32 layers to 4 and re-trained | Multilingual transcription; not trained for translation | MIT | Not published by OpenAI on this track[51] |
| [Parakeet](https://aiwiki.ai/wiki/parakeet) TDT 0.6B v3 | Aug 2025 | 600M FastConformer-TDT | 25 European | CC-BY-4.0 | 6.34, at an RTFx of 3,332.74[37] |
| [Canary](https://aiwiki.ai/wiki/canary)-Qwen-2.5B | Jul 2025 | 2.5B; FastConformer encoder plus Qwen3-1.7B decoder | English | CC-BY-4.0 | 5.63, at an RTFx of 418.28[32] |
| [Qwen3](https://aiwiki.ai/wiki/qwen3)-ASR-1.7B | Jan 2026 | 1.7B, built on the Qwen3-Omni stack | 52 languages and dialects | Apache 2.0 | 5.76 as published in Cohere's comparison[35] |
| [Cohere Transcribe](https://aiwiki.ai/wiki/cohere_transcribe) 03-2026 | Mar 2026 | 2B Fast-Conformer encoder, lightweight decoder | 14 | Apache 2.0 | 5.42[34] |
| IBM Granite Speech 4.1 2B | Apr 2026 | 2B; conformer encoder plus Granite 4.0 1B decoder | 6 | Apache 2.0 | 5.33, at an RTFx of 231.29[33] |
| [Voxtral](https://aiwiki.ai/wiki/voxtral) Mini Realtime | Feb 2026 | 4B streaming model | 13 | Apache 2.0 | Not on this track; Mistral reports about 4 percent WER on FLEURS[42] |
| Kyutai STT | 2025 | 1B English-French and 2.6B English streaming models | 2 | CC-BY-4.0 | 6.4 for the 2.6B English model, at an RTFx of 88.37[41] |
| Omnilingual ASR | Nov 2025 | 300M-7B, four model families | 1,600+ | Apache 2.0 | Not on this track; CER below 10 for 78 percent of languages[38][39] |

Three cautions apply to every row. First, the gap between the leading entries is now well under one WER point, which is small relative to the difference between test sets within the average and to the variation between demographic slices discussed later in this article. Second, several of these figures were published by a developer comparing its own model with competitors, and a self-reported comparison is weaker evidence than an independent run. Third, the models differ so much in language coverage, licensing, timestamp support, and throughput that a single accuracy column cannot rank them for any real deployment.

### Streaming and speech-to-speech systems

Streaming recognizers convert audio incrementally, and the design question is how much future audio the model is allowed to see before committing to a word. Kyutai's delayed streams modeling frames recognition and synthesis as the same problem with the delay applied to different streams: a text stream delayed behind audio gives ASR, and the reverse gives text-to-speech. The released speech-to-text models are a 1-billion-parameter English and French model with a 0.5 second delay and a semantic voice-activity detector that estimates whether the speaker has finished, and a 2.6-billion-parameter English model with a 2.5 second delay. Kyutai reports word-level timestamps and up to 400 concurrent real-time streams on a single H100.[41]

Mistral's Voxtral Realtime, announced on 4 February 2026 and released under Apache 2.0, makes the same delay explicit as a configurable parameter, advertised down to under 200 milliseconds, and states that at a 2.4 second delay the streaming model matches the accuracy of the company's batch transcription model. Mistral reports roughly 4 percent WER on FLEURS across 13 languages.[42] The delay parameter is the honest form of a streaming accuracy claim: any streaming WER quoted without its delay setting is incomplete.

Full-duplex speech-to-speech models remove the transcript from the interface entirely. [Moshi](https://aiwiki.ai/wiki/moshi), published by [Kyutai](https://aiwiki.ai/wiki/kyutai_labs) in September 2024, generates audio tokens for both sides of a conversation in parallel and predicts time-aligned text tokens as a prefix to the audio tokens, an arrangement the authors call Inner Monologue. The paper reports a theoretical latency of 160 milliseconds and about 200 milliseconds in practice, and the model handles overlapping speech and interruptions natively rather than through an external turn-taking rule.[40] Commercially, similar behavior is offered through the OpenAI Realtime API and Gemini Live, while conversational-voice vendors including [Cartesia](https://aiwiki.ai/wiki/cartesia), [Rime](https://aiwiki.ai/wiki/rime_ai), [Sesame CSM](https://aiwiki.ai/wiki/sesame_csm), [Hume AI](https://aiwiki.ai/wiki/hume_ai), and [Inworld AI](https://aiwiki.ai/wiki/inworld_ai) build products on the same idea.

Commercial streaming transcription remains a distinct market. [Deepgram](https://aiwiki.ai/wiki/deepgram), [AssemblyAI](https://aiwiki.ai/wiki/assemblyai), [Gladia](https://aiwiki.ai/wiki/gladia), [Speechmatics](https://aiwiki.ai/wiki/speechmatics), [iFlytek](https://aiwiki.ai/wiki/iflytek), and [ElevenLabs](https://aiwiki.ai/wiki/elevenlabs) all publish streaming and batch endpoints, and [NVIDIA Riva](https://aiwiki.ai/wiki/nvidia_riva), now documented as part of NVIDIA's Speech NIM microservices, packages GPU-accelerated recognition, synthesis, and translation for self-hosted deployment across cloud, data center, and embedded targets.[53] These services usually add capabilities that research checkpoints omit, including diarization, word timestamps, redaction, keyword boosting, and formatting.

## Decoding and post-processing

Training assigns probabilities, while decoding turns those probabilities into one or more token sequences. Greedy decoding chooses the locally best token or path. Beam search keeps several partial hypotheses, pruning those with low combined scores. A decoder may combine the model score with an external language model through shallow fusion, rescore complete candidates in a second pass, or apply a lexicon constraint.

Contextual biasing increases the probability of terms expected in a particular session, such as contact names, place names, or product vocabulary. Biasing should be evaluated for both recall and false positives. A list that is too strong can force an expected name when the speaker said something acoustically similar. The application also needs a fallback for terms not present in the bias list. In models with a language-model decoder the mechanism has changed shape: rather than reweighting a decoding graph, the bias terms may be supplied as a prompt or a keyword list, as in the keyword-biasing feature documented for Granite Speech 4.1.[33] The evaluation requirement is unchanged, because a prompt strong enough to insert a rare term is also strong enough to insert it when it was not spoken.

Raw ASR output often passes through:

- inverse text normalization, which converts spoken forms such as "twenty dollars" into a display form;
- capitalization and punctuation restoration;
- profanity or redaction policies;
- timestamp estimation;
- confidence estimation;
- speaker attribution; and
- formatting for captions, subtitles, notes, or commands.

These operations can change the visible transcript after recognition. A punctuation model can improve readability without changing lexical WER, while an incorrect normalizer can turn a correct spoken form into a wrong number. High-stakes evaluation should preserve and inspect both lexical output and display-form output. Several 2026 recognizers now perform punctuation, casing, and formatting inside the model rather than as a separate stage, which removes a pipeline component but also removes the place where a formatting rule could be inspected or overridden.

Confidence values are not self-interpreting probabilities. A model can be overconfident or underconfident, and calibration may shift across domains. If confidence triggers human review, rejection, or an automated action, the threshold should be selected on representative development data and audited after domain or model changes. Systems whose decoder is a general-purpose language model raise an additional problem, because the sequence probability reflects the text prior as well as the acoustic evidence, and a fluent invented sentence can carry a high score.

## Evaluation

The most common lexical metric is [Word Error Rate](https://aiwiki.ai/wiki/word_error_rate), or WER. A dynamic-programming alignment between a reference transcript and a hypothesis counts substitutions \(S\), deletions \(D\), and insertions \(I\). With \(N\) reference words:

```text
WER = (S + D + I) / N
```

NIST describes the same error categories and warns that scores from different collections may not be directly comparable when transcript quality, test selection, or word mapping differs.[24] Because insertions are not bounded by the number of reference words, WER can exceed 100 percent. An aggregate corpus WER should be computed from total errors divided by total reference words, not by averaging utterance percentages unless that alternative is explicitly intended.

Character error rate applies the same alignment idea to characters. It is useful where word segmentation is ambiguous or where character sequences are a natural unit, but it gives a different weighting to long and short words. Token error rate and phone error rate are also used when the system outputs those units. No one metric captures semantic severity: confusing "fifteen" with "fifty" may be more consequential than several harmless function-word errors.

Scoring requires a normalization protocol. Case, punctuation, hesitations, contractions, numbers, and spelling variants must be handled consistently. NIST's Speech Recognition Scoring Toolkit, SCTK, provides tools including SCLITE for alignment and scoring.[25] Publishing the scorer version, normalization rules, and reference conventions is part of a reproducible result.

Latency and efficiency are separate from recognition error:

- **Real-time factor** is processing time divided by audio duration under stated hardware and batching conditions. Its reciprocal, RTFx, is the figure most 2026 model cards report.
- **First-token latency** measures delay before useful partial output.
- **Endpoint latency** measures delay between the end of speech and a final result.
- **Revision behavior** measures how much a streaming partial transcript changes.
- **Memory, energy, and model size** affect whether the system fits a server, browser, or edge device.

Batch throughput can improve by processing many recordings together while making an individual request wait longer. A "real-time" claim is therefore incomplete without the latency statistic, hardware, concurrency, audio duration, and whether network time is included. The spread is large enough to matter: on the same leaderboard track, published RTFx values in 2026 range from a few hundred for models with language-model decoders to several thousand for a transducer model a quarter of their size.[32][33][37]

Evaluation data should be separated by speaker from training and adaptation data when speaker-independent performance is the goal. It should cover target microphones, noise, speaking styles, languages, accents, ages, and relevant speech impairments. Reported slices need enough samples to support conclusions, and an aggregate score should not hide a subgroup with substantially worse performance.

Benchmark contamination is another concern. Public test audio or transcripts can enter large web-scale training collections. A suspiciously strong result does not itself prove contamination, but dataset provenance, de-duplication, and evaluation on newly collected or private holdouts help establish that a system generalizes.

### Contamination and held-out evaluation

The concern is not hypothetical, and attaching a text language model to a speech encoder makes it worse. Tseng and colleagues showed in 2025 that substantial parts of the LibriSpeech and Common Voice evaluation sets appear in public language-model pretraining corpora, which is unsurprising for LibriSpeech in particular because its audio comes from LibriVox readings of public-domain books whose texts circulate widely in text collections. Comparing recognizers built on language models trained with and without that contamination, the authors found only small differences in error rate but significantly higher probabilities assigned to transcripts the language model had seen during its own pretraining. They recommend evaluating language-model-based speech systems on genuinely held-out data.[50]

The probability finding is the more useful one. A contaminated system may look only slightly better on WER while its confidence scores, N-best ordering, and rescoring behavior are all distorted, which matters for any downstream use that consumes uncertainty rather than the single best string. It also means a clean WER comparison between two systems can be misleading if only one of them has seen the reference text before. [Benchmark Contamination](https://aiwiki.ai/wiki/benchmark_contamination) is now a standard caveat for any speech result that relies on a pretrained text decoder, and the practical response is the same one the community reached for text models: reserve newly collected or private test audio, and report when a public set was used.

A related but distinct problem is overfitting through repeated use. LibriSpeech has been the default architecture benchmark for more than a decade, and its test-clean split is close to saturated, with several 2026 models reporting WER between 1.2 and 1.7 on it.[33][34][36] Differences at that scale are within the range of reference-transcript errors and normalization choices, so LibriSpeech test-clean no longer separates competent systems from each other. Harder splits, long-form audio, meeting corpora such as AMI, and financial or parliamentary speech such as Earnings-22 and VoxPopuli carry most of the discriminative signal in the current leaderboard average.

### What word error rate does not capture

WER counts edits, and every edit costs the same. That was a reasonable simplification when recognizers produced lowercase unpunctuated words and the reference used the same convention. It is a poor summary for a system whose decoder is a language model, because such a system produces display-ready text with its own defensible conventions: it may write "going to" where the speaker said "gonna," expand or contract a number, insert a comma, choose a different but correct spelling of a name, or lightly regularize a disfluent sentence. Each of these is counted as an error against a verbatim reference even though nothing about the meaning changed, while a single misheard drug name or account number counts exactly the same.

Parulekar and Jyothi make this argument directly in their EMNLP 2025 work on LASER, an LLM-based scoring rubric. They observe that standard metrics unfairly penalize morphological and syntactic variation that does not change sentence semantics, a problem that is sharpest for morphologically rich languages, and propose using a language model to align hypothesis and reference word by word and assign a graded penalty per error type. They report a 94 percent correlation with human annotation for Hindi using Gemini 2.5 Pro, find that Hindi examples in the prompt transfer usefully to Marathi, Kannada, and Malayalam, and show that a fine-tuned Llama 3 predicts the appropriate penalty type with about 89 percent accuracy.[48] Metrics of this kind are themselves model-dependent and inherit the judge model's biases, so they supplement WER rather than replacing it.

Several practical responses do not require an LLM judge:

- Report WER against both a verbatim reference and a normalized reference, so that formatting differences are visible separately from recognition errors.
- Report entity-level accuracy for the fields that carry consequence, such as names, dosages, amounts, dates, and identifiers, alongside the corpus WER.
- Add a semantic score. The Interspeech 2025 Speech Accessibility Project Challenge scored submissions on WER and a semantic score together, and the two rankings did not agree: 12 of 22 valid teams beat the Whisper large-v2 baseline on WER while 17 beat it on the semantic score.[55]
- For conversational systems, score the downstream action rather than the string.

Benchmarks for audio-capable language models measure something different again. MMAU pairs 10,000 audio clips with human-written questions across speech, environmental sound, and music, covering 27 skills and requiring reasoning rather than transcription; at publication the best evaluated systems scored around 53 percent.[49] A model can transcribe well and reason poorly about what it heard, or the reverse, so neither number substitutes for the other. The [SUPERB](https://aiwiki.ai/wiki/superb_benchmark) benchmark plays a comparable role for evaluating speech representations across multiple downstream tasks.

### Benchmark suites and leaderboards

The Open ASR Leaderboard, described in a 2025 paper by Srivastav and colleagues and updated through 2026, is the most widely cited public comparison. It evaluates 86 open-source and proprietary systems over 12 datasets in three tracks, English short-form, English long-form, and multilingual short-form, and standardizes both WER and RTFx so that accuracy and efficiency can be read together. Its dataset loaders and evaluation code are open, and it accepts community submissions across toolkits including ESPnet, NeMo, SpeechBrain, and Transformers.[31]

Its main structural finding is worth separating from any individual ranking. Conformer encoders with transformer decoders lead on average WER; CTC and token-and-duration transducer decoders lead on RTFx by a wide margin and are the better choice for long-form and batched work.[31] That is an architectural tradeoff, not a temporary state of the leaderboard, and it explains why a 600-million-parameter transducer remains in production use at organizations that could afford a 2-billion-parameter model with a language-model decoder.

Reading any leaderboard requires care about what the average hides. The English average is over eight test sets whose individual WERs, for a single 2026 model, can span from about 1.3 on LibriSpeech clean to about 11 on Earnings-22.[34] A model that is better on average may be worse on the set that resembles a given deployment. Ranking positions separated by a tenth of a point are not meaningful evidence of superiority for any particular use.

## Streaming and deployment

Offline recognition can wait for a complete utterance or recording and use future context throughout decoding. Streaming recognition must process audio incrementally and emit partial results before the stream ends. Causal encoders use only past frames. Limited-look-ahead systems also use a bounded amount of future audio. Chunked systems process blocks and cache left context.

There is no universal chunk size or latency threshold. The appropriate design depends on conversational turn-taking, caption readability, endpointing, compute, and network conditions. The streaming RNN-T study by Rao and colleagues demonstrates one architecture family that supports incremental processing,[22] while later work on streaming end-to-end ASR shows that endpoint penalties, minimum-WER training, and second-pass rescoring can change both recognition error and measured endpoint latency in a specific voice-search setup.[26]

Endpoint detection is often as important as encoder speed. Ending too early truncates a speaker; ending too late makes an assistant feel unresponsive. Pauses are not reliable sentence boundaries for every speaker or language, so endpoint rules should be evaluated on natural interaction data, including slow speech and speech impairments.

Deployment location changes the system boundary:

- A cloud service can pool accelerators and update models centrally, but it sends audio or derived data across a network and adds transport latency.
- An on-device or [Edge Computing](https://aiwiki.ai/wiki/edge_computing) system can keep audio local and work offline, but it faces stricter memory, energy, and thermal limits.
- A hybrid system can run wake-word or first-pass recognition locally and invoke a server for selected requests.

[Quantization](https://aiwiki.ai/wiki/quantization), pruning, distillation, caching, and smaller vocabularies can reduce resource use, but each change needs task-specific error and latency testing. Model-file size alone does not determine working memory or power consumption.

Privacy and security requirements apply to the whole pipeline. Audio can contain biometric, health, location, or confidential information even when the final transcript does not. A deployment should define retention, encryption, access control, logging, deletion, and whether recordings or corrections are reused for training. Human review adds another access path. Contextual bias lists can also reveal contacts or sensitive terms.

### Latency budgets for conversational systems

Conversational latency targets are set by human behavior rather than by engineering preference. Stivers and colleagues measured question-answer transitions in ten languages and found that every language showed a unimodal distribution of response offsets peaking within 200 milliseconds of the end of the question, with cross-language variation in the mean confined to a range of about 250 milliseconds.[54] A system that reliably takes a second to respond is outside the distribution people are used to, which is why perceived responsiveness, not transcription accuracy, is usually the first complaint about a voice product.

In a cascaded design the budget is spent across several stages in sequence: deciding that the user has stopped speaking, finalizing the transcript, generating the first token of a reply, synthesizing the first audio, and moving bytes across the network. Two properties of that arithmetic are worth stating plainly. The turn-taking decision is often the largest single term and the one least visible in benchmarks, because a recognizer scored offline never has to decide when a speaker has finished. And the terms are additive, so a design that is comfortable at each stage can still miss the target overall.

Three design responses are in current use. The first is to make endpointing semantic rather than acoustic: Kyutai ships a voice-activity detector that estimates the probability the user has finished speaking rather than waiting for a fixed silence duration, which removes a fixed pause from the budget.[41] The second is to expose the streaming delay as a tunable parameter and let the application choose its position on the accuracy curve, as Voxtral Realtime does with a delay configurable below 200 milliseconds and a stated accuracy match to batch transcription at 2.4 seconds.[42] The third is to collapse the chain entirely with a speech-to-speech model, which is where Moshi's reported 200 milliseconds in practice comes from: there is no separate recognition, generation, and synthesis stage to add up.[40]

Latency numbers should always be reported with their measurement point. First-partial latency, final-transcript latency after end of speech, and end-to-end time to first audio are three different quantities, and a figure quoted without hardware, concurrency, network path, and audio conditions is not reproducible.

### On-device and cloud

On-device recognition is now practical for general dictation and long-form transcription on consumer hardware, which was not true a few years ago. Apple's SpeechAnalyzer framework, introduced with iOS 26, coordinates on-device modules for transcription and voice activity detection and is documented as designed for sustained transcription over long recordings rather than short queries, replacing the previous API's session limits.[52] The same shift is visible in open models: distillation and decoder reduction produced Whisper large-v3-turbo in October 2024, which cuts the decoder from 32 layers to 4 for a total of 809 million parameters,[51] and Meta's Omnilingual ASR ships 300-million-parameter variants alongside its 7-billion-parameter model specifically for constrained hardware.[39] [Knowledge Distillation](https://aiwiki.ai/wiki/knowledge_distillation) and quantization are the usual levers, and both need re-measurement rather than an assumed accuracy cost.

The choice between local and remote execution is rarely made on accuracy alone. Local execution keeps audio on the device, works without connectivity, and removes per-request billing, at the cost of tighter memory and thermal limits, a fixed model that cannot be updated per request, and worse handling of rare vocabulary. Cloud execution pools accelerators, allows larger models and immediate updates, and supports diarization and other services that are impractical on a phone, at the cost of transport latency and a data-handling obligation. Regulated domains often force the decision: audio that cannot leave a jurisdiction cannot be sent to a shared endpoint, which is one reason self-hostable stacks such as NVIDIA's Speech NIM microservices exist alongside hosted APIs.[53]

Server-side deployment has its own tradeoff between latency and throughput. A batched offline pipeline maximizes RTFx and tolerates queueing; a conversational endpoint holds a session open per user and is bound by concurrency rather than aggregate throughput. Kyutai reports up to 400 concurrent real-time streams for its streaming recognizer on a single H100, which is the kind of figure that determines cost per session and does not appear on an accuracy leaderboard.[41]

### Cost

Hosted transcription is billed per minute of audio, with streaming priced above batch because it holds a session and cannot be packed as efficiently. Published list prices give the order of magnitude: as of July 2026 Mistral lists its batch transcription model at 0.003 US dollars per minute of audio and its realtime model at 0.006 dollars per minute, or roughly 0.18 and 0.36 dollars per hour.[42] Add-ons such as diarization, word-level timestamps, redaction, or domain modes are typically billed separately, so the effective price of a configured pipeline can be several times the headline rate.

Self-hosting shifts the question from price per minute to throughput per accelerator, which is why RTFx belongs beside WER in any procurement comparison. The published gap is large: a 600-million-parameter transducer reporting RTFx above 3,000 transcribes roughly an order of magnitude more audio per GPU-hour than a 2-billion-parameter model with a language-model decoder reporting RTFx in the low hundreds, for a WER difference of about one point on the same test suite.[33][37] For an archive of recorded audio that difference decides the compute bill; for a live agent handling one caller at a time it is close to irrelevant, because the constraint is concurrency and delay rather than aggregate throughput.

Voice agents are billed across a stack rather than a single model, since a cascaded design pays for recognition, generation, and synthesis on every turn, and a speech-to-speech model bills audio tokens in both directions. Comparing a per-minute transcription price against a per-token conversational price without modeling turn length and interruption rate will give the wrong answer.

## Robustness, inclusion, and failure modes

Background noise, reverberation, channel distortion, and overlapping speakers can all move input away from the training distribution. The CHiME-5 challenge was designed around distant, multi-microphone conversational speech in real homes and illustrates why clean close-talk results do not characterize far-field recognition.[27] Noise augmentation, beamforming, dereverberation, speech enhancement, multi-channel modeling, and domain adaptation are possible responses, but enhancement can also remove speech cues or introduce artifacts.

Accents and dialects are structured language varieties, not random noise. A 2017 study of YouTube automatic captions found accuracy differences across the gender and dialect groups in its sample and argued for sociolinguistically stratified validation.[28] A 2020 PNAS study evaluated five commercial ASR systems on two conversational-speech corpora and reported an average WER of 0.35 for Black speakers versus 0.19 for white speakers in that study.[29] Those figures describe the tested systems, data, and time period. They do not establish a permanent rate for a vendor or every speaker, but they demonstrate that aggregate performance can conceal substantial disparities.

Low-resource languages face more than a shortage of hours. Orthography may be unsettled, literacy and speech communities may not coincide, and available recordings may come from a narrow domain, region, or reading style. Community governance, consent, licensing, and decisions about scripts and normalization are part of building a useful resource. Common Voice demonstrates a crowdsourced approach,[20] while FLEURS provides parallel evaluation data across 102 languages.[21] The MMS work demonstrates much broader model coverage but also relies heavily on read religious texts,[23] and Omnilingual ASR extends coverage further while collecting new field recordings with local organizations for languages that public corpora did not reach.[38][39] These resources are complementary rather than interchangeable.

Code-switching combines acoustic and language-model challenges. A recognizer may choose the wrong language, normalize a borrowed word into the wrong script, or prefer a monolingual sequence. Test sets should preserve natural switch points and identify the languages and communities represented. Concatenating monolingual sentences can test engineering behavior, but it does not reproduce every property of spontaneous code-switching.

Names, numbers, abbreviations, and rare terms deserve targeted analysis. Their frequency is low, but their cost can be high in addresses, financial amounts, medication names, or legal records. A domain language model or bias phrase can help, yet an over-weighted prior can insert the expected term when it was not spoken.

### Measuring disparity

Measuring a disparity requires evaluation data annotated with the relevant speaker attributes, which most standard corpora lack. Meta's Fair-Speech dataset, published at Interspeech 2024, was assembled for this purpose: about 26,500 utterances of voice commands recorded by 593 paid participants in the United States, annotated with self-reported gender, age, ethnicity, socioeconomic background, and whether the speaker considers themselves a native English speaker.[57] Common Voice also collects optional self-reported demographic metadata, which makes its labeled subset usable for the same kind of slicing.[20]

A 2026 study by Ginjala, Fosler-Lussier, Myers, and Parthasarathy used both to test whether attaching a language-model decoder to a recognizer amplifies demographic bias. Evaluating nine systems across three architectural generations, a CTC model with no language model, three Whisper checkpoints with an implicit language model, and five systems with an explicit LLM decoder, over roughly 43,000 utterances and five demographic dimensions under clean and degraded audio, the authors report that LLM decoders did not amplify racial disparity as they had expected. Black and African American speakers had the highest error rates in every generation tested, which is a persistence finding rather than an amplification one. They also report that audio compression predicted accent fairness better than model scale did, and that Whisper large-v3 showed an insertion-rate spike to 9.62 percent on Indian-accented speech driven by repetition loops, while the explicit-LLM decoders stayed below 3.1 percent on the same material.[56] The insertion result connects fairness measurement to the hallucination behavior described below: the same decoder dynamics that invent text also concentrate the damage on speech that is furthest from the training distribution.

### Age and disordered speech

Recognizers trained mostly on adult speech transfer poorly to children. Children have higher fundamental and formant frequencies, more variable pitch and timing, and higher disfluency rates than adults, and they also change how they talk when addressing a device. A 2024 Scientific Reports study found that both adults and children used longer durations and higher pitch when speaking to a voice assistant than to a person, with children raising pitch more sharply after a staged recognition error, which the authors read as an adaptation learned from being misunderstood. The same paper summarizes prior findings that voice assistants respond correctly to only about half of queries from children aged 5 to 10.[60]

Disordered speech has been studied more systematically since the Speech Accessibility Project began collecting data at the University of Illinois Urbana-Champaign with a group of technology companies. The Interspeech 2025 challenge built on that corpus used more than 400 hours of speech and over 190,000 utterances from more than 500 speakers with Parkinson's disease, Down syndrome, amyotrophic lateral sclerosis, cerebral palsy, or the effects of stroke. Submissions were scored on WER and a semantic score. The best submission reached 8.11 percent WER with a semantic score of 88.44, and 12 of 22 valid teams beat the Whisper large-v2 baseline on WER while 17 beat it on the semantic score.[55][61]

Two things follow from that result. Substantial improvement on disordered speech is achievable with targeted data, which means a large gap is a data and evaluation choice rather than an inherent limit. And an 8 percent WER on a challenge set is not the same as usable performance for an individual speaker, because severity varies widely within every diagnosis and per-speaker adaptation matters more here than in most other conditions. Any product claim about accessibility should be supported by per-speaker results, not a corpus average.

### Hallucinated transcription

Neural sequence models can emit fluent text that is unsupported by the audio, especially in difficult or low-information segments. A FAccT 2024 study examined Whisper transcripts and found that roughly 1 percent of the evaluated audio segments contained an entire hallucinated phrase or sentence; the study also found a higher rate for speech with longer non-vocal durations, including recordings from speakers with aphasia.[30] This is a bounded result for that study and model evaluation, not a universal hallucination rate. It is enough to show that fluent output and confidence should not be treated as evidence that words were present.

The failure has a structural explanation. An autoregressive decoder trained to produce fluent text will produce fluent text when the acoustic evidence is weak, because nothing in the objective requires it to abstain. Silence, music, breathing, and background noise are all inputs the decoder was never rewarded for refusing. Baranski and colleagues examined this directly at ICASSP 2025, cataloguing recurring hallucinations that non-speech audio triggers in Whisper and showing that a post-processing filter built from the observed set of hallucinated strings reduces WER and acts as a partial safeguard.[58] Cohere's 2026 model card states the same tendency for its own system in plain terms, warning that the model "is eager to transcribe, even non-speech sounds" and recommending preprocessing.[34]

The documented consequences have been most serious in clinical settings. An Associated Press investigation published in October 2024 reported that a Whisper-based clinical documentation tool from Nabla was in use by more than 30,000 clinicians and 40 health systems and had been used for an estimated 7 million medical visits. The report described one machine-learning engineer finding hallucinations in about half of more than 100 hours of Whisper transcripts he examined, and another finding them in nearly all of 26,000 transcripts, with invented content including nonexistent medications and remarks the speakers never made. It also reported that the tool deleted the original audio for stated data-safety reasons, which removes the only means of checking a suspect transcript against what was actually said.[59] Vendors have responded: OpenAI's December 2025 audio-model notes claim roughly 90 percent fewer hallucinations than Whisper v2 on the company's own noise testing.[45] That is a self-reported improvement on an internal evaluation, not an independent measurement, and it is a reduction rather than an elimination.

Deleting source audio is the specific practice worth singling out, because it converts a recoverable transcription error into an unfalsifiable record. Whatever the retention policy, a clinical, legal, or safety-relevant transcript that cannot be checked against its audio should not be treated as a record of what was said.

High-stakes use therefore needs controls beyond average WER. Depending on the application, controls can include retaining source audio, displaying timestamps, flagging low-confidence spans, preserving alternative hypotheses, requiring human confirmation for critical fields, and preventing unreviewed transcripts from triggering irreversible actions. A recognizer should not be represented as a diagnostic, legal, or factual authority merely because its output is grammatical.

## Applications

Speech recognition supports captioning, dictation, meeting and call transcription, media indexing, voice search, hands-free interfaces, contact-center analytics, accessibility tools, and data entry. It can provide text to downstream translation, summarization, retrieval, or dialogue systems.

Application requirements vary:

- **Captions** need low delay, readable segmentation, stable partial text, and correct speaker or sound-event handling.
- **Dictation** benefits from domain vocabulary, editing commands, punctuation, and document formatting.
- **Meeting transcription** adds far-field acoustics, overlap, diarization, and long-context challenges.
- **Voice commands** may care more about correctly recognizing a small set of actionable slots than about full verbatim WER.
- **Search and indexing** can use lattices or confidence-weighted alternatives instead of one transcript.
- **Archival transcription** can trade latency for multiple decoding passes and human correction.

Product categories have consolidated around these requirements. Meeting and call tools such as [Otter.ai](https://aiwiki.ai/wiki/otter_ai), [Fireflies.ai](https://aiwiki.ai/wiki/fireflies_ai), and [Descript](https://aiwiki.ai/wiki/descript) combine recognition with diarization, search, and editing; local dictation utilities such as [Superwhisper](https://aiwiki.ai/wiki/superwhisper) run open recognizers on a laptop; [Krisp AI](https://aiwiki.ai/wiki/krisp_ai) sits upstream of recognition as noise suppression; and [Apple Intelligence](https://aiwiki.ai/wiki/apple_intelligence) and comparable platform features embed transcription in the operating system. The fastest-growing category is the [AI Voice Agent](https://aiwiki.ai/wiki/ai_voice_agent), which combines recognition, a language model, and [Text-to-Speech](https://aiwiki.ai/wiki/text_to_speech) in a loop and inherits every constraint discussed above at once. The older [Voice Assistant](https://aiwiki.ai/wiki/voice_assistant) products remain the largest deployed base by user count.

The transcript is often an intermediate representation. If the final task is intent classification or entity extraction, both transcription metrics and task metrics should be reported. A lower WER does not guarantee a lower error rate on every downstream decision because different words have different consequences. In systems built on omni models the transcript may not exist as a separate artifact at all, which removes the most convenient debugging surface and makes task-level evaluation mandatory rather than optional.

## Building and reporting an ASR system

A defensible workflow begins with a written task specification:

1. Define languages, domains, microphones, channels, speaking styles, and output conventions.
2. Establish consent, licensing, retention, and permitted uses for every data source.
3. Split training, development, and test data to prevent speaker, recording, or near-duplicate leakage.
4. Choose output units and normalization rules before comparing systems.
5. Select a modular, CTC, transducer, attention-based, or language-model-decoder design according to evidence, latency, and maintenance needs rather than a universal ranking.
6. Train or adapt on representative data and keep the final test set isolated.
7. Tune decoding, biasing, endpointing, and confidence thresholds on development data.
8. Evaluate lexical errors, critical entities, subgroup slices, robustness, latency, compute, and downstream outcomes.
9. Inspect errors manually, including fluent insertions and failures around silence or overlap.
10. Monitor drift after deployment and rerun the same controlled evaluation after data, model, decoder, or normalization changes.

A reproducible report should identify:

- the model architecture and exact checkpoint;
- the training and adaptation datasets, licenses, filters, and split rules;
- audio sampling, segmentation, features, and augmentation;
- output vocabulary and tokenization procedure;
- optimization schedule, random seeds, and stopping criterion;
- external language models, lexicons, bias lists, prompts, and decoding weights;
- transcript normalization and scoring software;
- hardware, precision, batching, concurrency, and latency measurement protocol;
- overall and sliced error metrics, with RTFx alongside WER;
- known failure modes and intended uses; and
- whether test data may overlap web-scale pretraining sources, including the text corpora used to train a language-model decoder.

These details matter because "the same model" can produce different results after a tokenizer, decoding weight, endpointing rule, prompt, or text-normalization change. Reporting only a model name and WER is not enough to reproduce or safely interpret a result, and it has become less sufficient as decoders have acquired stronger text priors and more configurable behavior.

## References

1. K. H. Davis, R. Biddulph, and S. Balashek, "Automatic Recognition of Spoken Digits," *Journal of the Acoustical Society of America*, 24(6), 637-642, 1952. https://doi.org/10.1121/1.1906946

2. IBM, "Voice recognition." https://www.ibm.com/history/voice-recognition

3. F. Jelinek, "Continuous Speech Recognition by Statistical Methods," *Proceedings of the IEEE*, 64(4), 532-556, 1976. https://research.ibm.com/publications/continuous-speech-recognition-by-statistical-methods

4. L. R. Rabiner, "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition," *Proceedings of the IEEE*, 77(2), 257-286, 1989. https://doi.org/10.1109/5.18626

5. Carnegie Mellon University, "The Sphinx History." https://www.cs.cmu.edu/~rsingh/homepage/sphinx_history.html

6. M. Mohri, F. Pereira, and M. Riley, "Weighted Finite-State Transducers in Speech Recognition," *Computer Speech & Language*, 16(1), 69-88, 2002. https://doi.org/10.1006/csla.2001.0184

7. University of Cambridge, "HTK License." https://htk.eng.cam.ac.uk/docs/license.shtml

8. D. Povey et al., "The Kaldi Speech Recognition Toolkit," *IEEE Workshop on Automatic Speech Recognition and Understanding*, 2011. https://kaldi-asr.org/doc/about.html

9. G. Hinton et al., "Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups," *IEEE Signal Processing Magazine*, 29(6), 82-97, 2012. https://doi.org/10.1109/MSP.2012.2205597

10. A. Graves, S. Fernandez, F. Gomez, and J. Schmidhuber, "Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks," *Proceedings of the 23rd International Conference on Machine Learning*, 369-376, 2006. https://doi.org/10.1145/1143844.1143891

11. A. Graves and N. Jaitly, "Towards End-to-End Speech Recognition with Recurrent Neural Networks," *Proceedings of Machine Learning Research*, 32(2), 1764-1772, 2014. https://proceedings.mlr.press/v32/graves14.html

12. W. Chan, N. Jaitly, Q. Le, and O. Vinyals, "Listen, Attend and Spell: A Neural Network for Large Vocabulary Conversational Speech Recognition," *IEEE ICASSP*, 4960-4964, 2016. https://doi.org/10.1109/ICASSP.2016.7472621

13. A. Graves, "Sequence Transduction with Recurrent Neural Networks," 2012. https://arxiv.org/abs/1211.3711

14. A. Gulati et al., "Conformer: Convolution-augmented Transformer for Speech Recognition," *Interspeech 2020*, 5036-5040. https://www.isca-archive.org/interspeech_2020/gulati20_interspeech.html

15. A. Baevski, H. Zhou, A. Mohamed, and M. Auli, "wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations," *Advances in Neural Information Processing Systems 33*, 2020. https://proceedings.neurips.cc/paper/2020/hash/92d1e1eb1cd6f9fba3227870bb6d7f07-Abstract.html

16. W.-N. Hsu et al., "HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units," *IEEE/ACM Transactions on Audio, Speech, and Language Processing*, 29, 3451-3460, 2021. https://doi.org/10.1109/TASLP.2021.3122291

17. A. Radford et al., "Robust Speech Recognition via Large-Scale Weak Supervision," *Proceedings of Machine Learning Research*, 202, 28492-28518, 2023. https://proceedings.mlr.press/v202/radford23a.html

18. D. S. Park et al., "SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition," *Interspeech 2019*, 2613-2617. https://www.isca-archive.org/interspeech_2019/park19e_interspeech.html

19. V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, "LibriSpeech: An ASR Corpus Based on Public Domain Audio Books," *IEEE ICASSP*, 5206-5210, 2015. https://doi.org/10.1109/ICASSP.2015.7178964

20. R. Ardila et al., "Common Voice: A Massively-Multilingual Speech Corpus," *Proceedings of LREC 2020*, 4218-4222. https://aclanthology.org/2020.lrec-1.520/

21. A. Conneau et al., "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech," *2022 IEEE Spoken Language Technology Workshop*, 798-805. https://doi.org/10.1109/SLT54892.2023.10023141

22. K. Rao, H. Sak, and R. Prabhavalkar, "Exploring Architectures, Data and Units for Streaming End-to-End Speech Recognition with RNN-Transducer," *2017 IEEE Automatic Speech Recognition and Understanding Workshop*, 193-199. https://research.google/pubs/exploring-architectures-data-and-units-for-streaming-end-to-end-speech-recognition-with-rnn-transducer/

23. V. Pratap et al., "Scaling Speech Technology to 1,000+ Languages," *Journal of Machine Learning Research*, 25(97), 1-52, 2024. https://www.jmlr.org/papers/v25/23-1318.html

24. National Institute of Standards and Technology, "ASR Metrics." https://trec.nist.gov/pubs/trec9/sdrt9_slides/tsld017.htm

25. National Institute of Standards and Technology, "Tools: Speech Recognition Scoring Toolkit." https://www.nist.gov/itl/iad/mltg/tools

26. B. Li et al., "Towards Fast and Accurate Streaming End-to-End ASR," *IEEE ICASSP*, 6069-6073, 2020. https://research.google/pubs/towards-fast-and-accurate-streaming-end-to-end-asr/

27. J. Barker et al., "The Fifth 'CHiME' Speech Separation and Recognition Challenge: Dataset, Task and Baselines," *Interspeech 2018*, 1561-1565. https://www.isca-archive.org/interspeech_2018/barker18_interspeech.html

28. R. Tatman, "Gender and Dialect Bias in YouTube's Automatic Captions," *Proceedings of the First ACL Workshop on Ethics in Natural Language Processing*, 53-59, 2017. https://aclanthology.org/W17-1606/

29. A. Koenecke et al., "Racial Disparities in Automated Speech Recognition," *Proceedings of the National Academy of Sciences*, 117(14), 7684-7689, 2020. https://doi.org/10.1073/pnas.1915768117

30. A. Koenecke, A. S. G. Choi, K. X. Mei, H. Schellmann, and M. Sloane, "Careless Whisper: Speech-to-Text Hallucination Harms," *Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency*, 1672-1681. https://doi.org/10.1145/3630106.3658996

31. V. Srivastav, S. Zheng, E. Bezzam, E. Le Bihan, N. R. Koluguri, P. Zelasko, S. Majumdar, A. Moumen, and S. Gandhi, "Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation," arXiv:2510.06961, October 2025 (revised March 2026). https://arxiv.org/abs/2510.06961

32. NVIDIA, "canary-qwen-2.5b" model card, Hugging Face, July 2025. https://huggingface.co/nvidia/canary-qwen-2.5b

33. IBM, "granite-speech-4.1-2b" model card, Hugging Face, April 2026. https://huggingface.co/ibm-granite/granite-speech-4.1-2b

34. Cohere Labs, "cohere-transcribe-03-2026" model card, Hugging Face, March 2026. https://huggingface.co/CohereLabs/cohere-transcribe-03-2026

35. Cohere Labs, "Introducing Cohere-transcribe: state-of-the-art speech recognition," Hugging Face blog, March 2026. https://huggingface.co/blog/CohereLabs/cohere-transcribe-03-2026-release

36. Qwen team, Alibaba Cloud, "Qwen3-ASR" repository, January 2026. https://github.com/QwenLM/Qwen3-ASR

37. NVIDIA, "parakeet-tdt-0.6b-v3" model card, Hugging Face, August 2025. https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3

38. Omnilingual ASR team, "Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages," arXiv:2511.09690, November 2025. https://arxiv.org/abs/2511.09690

39. Meta (Facebook Research), "omnilingual-asr" repository. https://github.com/facebookresearch/omnilingual-asr

40. A. Defossez, L. Mazare, M. Orsini, A. Royer, P. Perez, H. Jegou, E. Grave, and N. Zeghidour, "Moshi: a speech-text foundation model for real-time dialogue," arXiv:2410.00037, September 2024. https://arxiv.org/abs/2410.00037

41. Kyutai, "Kyutai STT" (https://kyutai.org/stt/) and the "stt-2.6b-en" model card, Hugging Face. https://huggingface.co/kyutai/stt-2.6b-en

42. Mistral AI, "Voxtral transcribes at the speed of sound," 4 February 2026. https://mistral.ai/news/voxtral-transcribe-2/

43. OpenAI, "Introducing next-generation audio models in the API," 20 March 2025. https://openai.com/index/introducing-our-next-generation-audio-models/

44. OpenAI, "Introducing gpt-realtime and Realtime API updates for production voice agents," 28 August 2025. https://openai.com/index/introducing-gpt-realtime/

45. OpenAI, "Updates for developers building with voice," developer blog, December 2025. https://developers.openai.com/blog/updates-audio-models

46. Google, "Live API overview," Gemini API documentation. https://ai.google.dev/gemini-api/docs/live-api

47. Qwen team, Alibaba Cloud, "Qwen3-Omni" repository. https://github.com/QwenLM/Qwen3-Omni

48. A. Parulekar and P. Jyothi, "LASER: An LLM-based ASR Scoring and Evaluation Rubric," *Proceedings of EMNLP 2025*; arXiv:2510.07437. https://arxiv.org/abs/2510.07437

49. S. Sakshi, U. Tyagi, S. Kumar, A. Seth, R. Selvakumar, O. Nieto, R. Duraiswami, S. Ghosh, and D. Manocha, "MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark," arXiv:2410.19168, October 2024. https://arxiv.org/abs/2410.19168

50. Y. Tseng, T. Parcollet, R. van Dalen, S. Zhang, and S. Bhattacharya, "Evaluation of LLMs in Speech is Often Flawed: Test Set Contamination in Large Language Models for Speech Recognition," arXiv:2505.22251, May 2025. https://arxiv.org/abs/2505.22251

51. OpenAI, "turbo model release," openai/whisper discussion #2363, October 2024. https://github.com/openai/whisper/discussions/2363

52. Apple, "SpeechAnalyzer," Speech framework documentation. https://developer.apple.com/documentation/speech/speechanalyzer

53. NVIDIA, "NVIDIA Speech NIM Microservices" documentation. https://docs.nvidia.com/nim/speech/latest/index.html

54. T. Stivers, N. J. Enfield, P. Brown, C. Englert, M. Hayashi, T. Heinemann, G. Hoymann, F. Rossano, J. P. de Ruiter, K.-E. Yoon, and S. C. Levinson, "Universals and Cultural Variation in Turn-Taking in Conversation," *Proceedings of the National Academy of Sciences*, 106(26), 10587-10592, 2009. https://doi.org/10.1073/pnas.0903616106

55. X. Zheng et al., "The Interspeech 2025 Speech Accessibility Project Challenge," arXiv:2507.22047, July 2025. https://arxiv.org/abs/2507.22047

56. A. Ginjala, E. Fosler-Lussier, C. Myers, and S. Parthasarathy, "Do LLM Decoders Listen Fairly? Benchmarking How Language Model Priors Shape Bias in Speech Recognition," arXiv:2604.21276, April 2026. https://arxiv.org/abs/2604.21276

57. I.-E. Veliche, Z. Huang, V. A. Kochaniyan, F. Peng, O. Kalinli, and M. L. Seltzer, "Towards Measuring Fairness in Speech Recognition: Fair-Speech Dataset," *Interspeech 2024*. https://www.isca-archive.org/interspeech_2024/veliche24_interspeech.html

58. M. Baranski, J. Jasinski, J. Bartolewska, S. Kacprzak, M. Witkowski, and K. Kowalczyk, "Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio," *IEEE ICASSP 2025*; arXiv:2501.11378. https://arxiv.org/abs/2501.11378

59. G. Burke and H. Schellmann, "Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said," Associated Press, 28 October 2024. https://www.columbian.com/news/2024/oct/28/researchers-say-an-ai-powered-transcription-tool-used-in-hospitals-invents-things-no-one-ever-said/

60. M. Cohn, S. Barreda, K. Graf Estes, Z. Yu, and G. Zellou, "Children and Adults Produce Distinct Technology- and Human-Directed Speech," *Scientific Reports*, 14, 2024. https://www.nature.com/articles/s41598-024-66313-5

61. Beckman Institute, University of Illinois Urbana-Champaign, "Speech Accessibility Project." https://speechaccessibilityproject.beckman.illinois.edu/