NVIDIA Parakeet

Parakeet is a family of open automatic speech recognition (ASR) models developed by NVIDIA as part of the NeMo conversational AI toolkit. The models transcribe spoken English (and, from 2025, additional languages) by pairing an efficient FastConformer encoder with transducer-style or CTC decoders. Parakeet became widely known for topping the Hugging Face Open ASR Leaderboard in 2024 and again in 2025 while running an order of magnitude faster than the systems it displaced. The earliest Parakeet checkpoints were developed jointly by the NVIDIA NeMo and Suno.ai teams and published on Hugging Face at the end of 2023 under a permissive Creative Commons license.^[1][3]

The name Parakeet is used for the transducer and CTC branch of NVIDIA's NeMo speech models, distinguishing it from Canary, the companion family of attention-based encoder-decoder models that additionally performs speech translation. Both families share the same FastConformer encoder but differ in their decoders and in the breadth of tasks they support. Since late 2025 NVIDIA has marketed its speech models collectively as Nemotron Speech and has moved production deployment on x86 servers from the Riva SDK to NIM microservices, changes that affect how Parakeet is packaged rather than what it is.^[17][19]

The accuracy picture has also changed. On the Open ASR Leaderboard results file dated July 31, 2026, the strongest Parakeet entry (parakeet-tdt-0.6b-v2) ranked 27th of 74 systems by average word error rate, well behind a cluster of newer speech-LLM hybrids and proprietary APIs. Its speed advantage was intact: the seven fastest systems on that leaderboard, measured by inverse real-time factor, were all Parakeet checkpoints.^[12]

Background and positioning

Open ASR systems are typically benchmarked on the Hugging Face Open ASR Leaderboard, which ranks models by two complementary measures: the average word error rate (WER) across a suite of English test sets such as LibriSpeech, AMI, Earnings-22, GigaSpeech, SPGISpeech, TED-LIUM and VoxPopuli, and the inverse real-time factor (RTFx), which measures transcription speed as the ratio of audio duration to processing time.^[10] A higher RTFx means more hours of audio can be transcribed per hour of compute. Parakeet was engineered to win on both axes simultaneously, rather than trading accuracy for speed.

The leaderboard's own methodology paper, published in October 2025, describes it as "a reproducible benchmarking platform with community contributions from academia and industry" comparing 86 open-source and proprietary systems across 12 datasets, with English short-form, English long-form and multilingual short-form tracks. Its headline finding is a direct statement of the trade-off Parakeet sits on: "Conformer-based encoders paired with transformer-based decoders achieve the best average WER, while connectionist temporal classification (CTC) and token-and-duration transducer (TDT) decoders offer superior RTFx, making them better suited for long-form and batched processing."^[13]

The four initial Parakeet models, using CTC and RNN-T decoders, were published in late December 2023, followed by a TDT variant in January 2024, and they quickly reached the top of the leaderboard.^[1][2][3] In March 2024 NVIDIA wrote that the Parakeet family and the Canary model "currently top the Hugging Face Open ASR Leaderboard."^[9] The models were trained to be resilient to non-speech audio such as music and silence, and to generalize across a wide range of accents, dialects, vocal ranges and noise conditions.^[1]

FastConformer architecture

All Parakeet models use a FastConformer encoder, an optimized variant of the Conformer architecture introduced by NVIDIA in the paper "Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition."^[7] Compared with the standard Conformer, FastConformer adds an additional 2x convolutional subsampling stage so that the input is downsampled by a factor of 8 overall, uses depthwise-separable convolutions in the subsampling layers, reduces the number of channels in those layers from 512 to 256, and shrinks the convolution kernel size inside the Conformer blocks from 31 to 9. These changes make the encoder roughly 2.8x faster than the original Conformer while preserving accuracy, and they allow the architecture to scale to roughly a billion parameters without structural changes.^[7]

FastConformer also supports limited (local) context attention, which lets a single model transcribe very long recordings without running out of memory. NVIDIA reported that the architecture could process up to 11 hours of speech in one pass on an 80 GB NVIDIA A100 GPU using local attention.^[7] The April 2024 Parakeet blog gives per-model figures at batch size 1: 14 hours for the 120M model, 13 hours for the 0.6B model and 12.5 hours for the 1.1B model.^[1] The 2025 checkpoints report different limits because they were trained with full attention instead: parakeet-tdt-0.6b-v2 handles audio up to 24 minutes in a single pass, and parakeet-tdt-0.6b-v3 handles 24 minutes with full attention or about three hours with local attention on an 80 GB A100.^[5][6]

A separate line of work adapts the same encoder for streaming. Cache-aware FastConformer variants, described in "Stateful Conformer with Cache-based Inference for Streaming Automatic Speech Recognition," keep encoder state between chunks so that left context is not recomputed on every step. This is the basis for NVIDIA's low-latency speech models, including parakeet_realtime_eou_120m-v1 and the Nemotron Speech streaming checkpoints.^[27]

Decoder variants

Parakeet's distinguishing feature is that the same FastConformer encoder is offered with several different decoders, each making a different trade-off between accuracy and inference speed.

• CTC (Connectionist Temporal Classification). A non-autoregressive softmax prediction head. Because each output does not depend on previous outputs, CTC decoding is highly parallel and very fast, at some cost in accuracy relative to transducer decoders.
• RNN-T (RNN-Transducer, RNNT). Adds an LSTM prediction network and a joint network to the encoder, producing an autoregressive model in which each token depends on the previously emitted tokens. This generally improves accuracy over CTC.
• TDT (Token-and-Duration Transducer). A refined transducer objective introduced by NVIDIA in the paper "Efficient Sequence Transduction by Jointly Predicting Tokens and Durations." A TDT model's joint network produces two independently normalized distributions: one over output tokens and one over durations, where the duration is the number of input frames the emitted token spans. By predicting durations, the decoder can skip ahead over blank frames during inference instead of advancing one frame at a time. The paper reported that TDT models achieved better accuracy and up to 2.82x faster inference than conventional transducers.^[8]

NeMo also ships hybrid TDT-CTC checkpoints (parakeet-tdt_ctc-110m, parakeet-tdt_ctc-1.1b and a Japanese parakeet-tdt_ctc-0.6b-ja) that combine a TDT decoder and a CTC decoder on a shared encoder, letting a single model serve either decoding path.

A fourth mode arrived in 2026. parakeet-unified-en-0.6b trains one FastConformer encoder jointly in offline and streaming modes and attaches a single RNN-T decoder, so that the same weights serve batch transcription and live captioning down to 160 ms of latency. The technique, mode-consistency regularization for RNN-T, is described in an April 2026 paper from the NeMo team.^[14][31]

The decoder distinction is durable enough that Hugging Face Transformers encodes it in the class hierarchy: native Parakeet support, contributed by NVIDIA and Hugging Face engineers, exposes a shared ParakeetEncoder alongside ParakeetForCTC, ParakeetForRNNT and ParakeetForTDT, with the decoder type controlling how timestamps are emitted.^[22]

The original model family

The first Parakeet release, jointly developed by the NVIDIA NeMo and Suno.ai teams, consisted of English-only models at 0.6 billion and 1.1 billion parameters. The Hugging Face model cards for these checkpoints state explicitly that they were "jointly developed by NVIDIA NeMo and Suno.ai teams."^[3][4] They were trained on 64,000 hours of English speech assembled by the two teams, made up of a private subset of 40,000 hours plus 24,000 hours from public datasets including LibriSpeech, the Fisher Corpus, Switchboard-1, WSJ-0 and WSJ-1, the National Speech Corpus, VCTK, VoxPopuli, Europarl-ASR, a 2,000-hour subset of Multilingual LibriSpeech, Mozilla Common Voice v7.0 and a 12,000-hour subset of People's Speech.^[3][4] These models transcribe into the lower-case English alphabet only: automatic punctuation and capitalization did not arrive until the 2025 checkpoints.^[3]

The Token-and-Duration Transducer model, Parakeet-TDT-1.1B, was highlighted as the most accurate of the family. NVIDIA reported that it was the first model to achieve an average WER below 7.0 on the Open ASR Leaderboard, that it ran 64% faster than the previously best Parakeet model (Parakeet-RNNT-1.1B), and that its real-time factor was 40% better than Parakeet-RNNT-0.6B despite being roughly twice the size.^[2] All of these checkpoints were released under the CC-BY-4.0 license, which permits commercial use.^[3][4]

The table below gives current leaderboard figures rather than the launch-era numbers, which were measured on different hardware and against a smaller field. "Cleaned" and "original" are the leaderboard's two English averages, computed over cleaned and uncleaned versions of the AMI, GigaSpeech and VoxPopuli references.

Scroll sideways for more →

Figures are from the Open ASR Leaderboard English short-form results file dated July 31, 2026.^[12]

NVIDIA has separately published throughput figures measured outside the leaderboard harness, using additional algorithmic and CUDA-level optimizations: 1,336 hours of audio transcribed per hour of real time for Parakeet-CTC-1.1B, 1,212 for Parakeet-TDT-1.1B and 1,120 for Parakeet-RNNT-1.1B.^[9] Numbers of this kind are not comparable across sources, because RTFx varies with GPU, batch size and the length of the audio clips being fed in. NVIDIA's own model cards carry the caveat directly.^[5]

Parakeet-TDT-0.6B-v2

In May 2025 NVIDIA released Parakeet-TDT-0.6B-v2, a 600-million-parameter FastConformer-TDT model for English transcription that topped the Hugging Face Open ASR Leaderboard at launch.^[5][11] NVIDIA reported an average WER of 6.05% across the leaderboard's test sets and an RTFx of 3,380 at batch size 128, equivalent to transcribing roughly 60 minutes of audio in about one second on GPU-accelerated hardware.^[5] A June 2025 NVIDIA post described the model as "#1 on the Hugging Face ASR leaderboard" with an "industry-best 6.05% word error rate."^[32] Beyond raw transcription, the model produces automatic punctuation and capitalization and accurate word-level timestamps, and it was tuned to handle spoken numbers and song-lyric transcription robustly.^[5]

Parakeet-TDT-0.6B-v2 was trained on the Granary dataset, a large open corpus of about 120,000 hours of English audio made up of roughly 10,000 hours of human-transcribed speech from NeMo ASR Set 3.0 and about 110,000 hours of pseudo-labeled data drawn from YouTube-Commons, YODAS and LibriLight.^[5] Like the rest of the family it was released under the CC-BY-4.0 license, making it fully usable in commercial products, and it runs through the NeMo toolkit.

It remains the most downloaded Parakeet checkpoint by a wide margin, with about 11.2 million cumulative downloads on Hugging Face as of August 1, 2026.^[5]

Parakeet-TDT-0.6B-v3 and the multilingual expansion

In August 2025 NVIDIA released Parakeet-TDT-0.6B-v3, extending the same 600M-parameter FastConformer-TDT design to 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian.^[6] The model detects the language of the audio itself and transcribes without a prompt.^[6]

Training used the multilingual Granary corpus: 660,000 hours of pseudo-labeled data plus roughly 10,000 hours of human-transcribed material from NeMo ASR Set 3.0, with a first stage of 150,000 steps on 128 A100 GPUs from a CTC multilingual checkpoint, then 5,000 steps of fine-tuning on 4 A100s using about 7,500 hours of high-quality human transcripts.^[6] Granary itself is documented in an Interspeech 2025 paper, "Granary: Speech Recognition and Translation Dataset in 25 European Languages," which reports that models trained on its pseudo-labeled data reach comparable quality using roughly 50% less data than unfiltered alternatives.^[23]

On the English leaderboard suite the v3 model card reports an average WER of 6.34%.^[6] It retains punctuation, capitalization, and word- and segment-level timestamps, and continues to support long-form audio through local attention.

Scroll sideways for more →

The leaderboard and model card numbers differ because the leaderboard re-runs every system on its own harness and its reference-cleaning rules changed in July 2026; the ordering of the two models is the same either way.^[12]

The wider family, 2024 to 2026

NVIDIA has continued to ship Parakeet-named checkpoints well past the flagship English models, mostly to cover specific languages, latency targets or conversational settings.

Scroll sideways for more →

Publication months are the Hugging Face repository creation dates unless the model card gives an explicit release date.^{[3][4][5][6][14][27][28][29][30]}

Three of these are worth separating out because they solve problems the flagship models do not.

End-of-utterance detection. parakeet_realtime_eou_120m-v1 is a 120M cache-aware streaming model with 17 encoder layers that emits an <EOU> token at the end of each utterance, so a downstream voice agent knows when the speaker has stopped without a separate turn-detection component. It runs at 80 ms to 160 ms latency and deliberately omits punctuation and capitalization.^[27]

Overlapping speakers. multitalker-parakeet-streaming-0.6b-v1 injects learnable speaker kernels into the pre-encode layer of the FastConformer encoder, driven by diarization output rather than by speaker enrollment audio or embeddings. NVIDIA runs one model instance per speaker, which costs compute proportional to the number of talkers but handles fully overlapped speech in both offline and streaming settings.^[28]

Unified latency. parakeet-unified-en-0.6b, trained on roughly 250,000 hours of US English from Granary, replaces the awkward choice between an offline model and a streaming one. NVIDIA's own comparison shows why the choice mattered: parakeet-tdt-0.6b-v2 degrades from 6.04% WER offline to 22.83% at a 1.12-second chunk and 95.12% at 400 ms, because it was never trained to stream, while the unified model holds 5.91% offline and 6.70% at 400 ms. Below 160 ms NVIDIA recommends a dedicated streaming model instead.^[14]

Nemotron Speech and the naming question

During 2025 and 2026 NVIDIA rebranded its speech portfolio. The developer page that used to front the Riva SDK is now headed "NVIDIA Nemotron Speech for Developers" and opens: "NVIDIA Nemotron Speech is a family of GPU-accelerated open models for automatic speech recognition (ASR), text-to-speech (TTS), neural machine translation (NMT), and speech-to-speech across ~40 languages. Deployed through the NVIDIA Riva library for optimized inference, the speech models are fully customizable and run across all clouds, in data centers, at the edge, and in embedded devices."^[17] The former nvidia.com product pages for Riva and Riva Enterprise now redirect to that page.

Parakeet has not been renamed. The evidence points to Nemotron Speech being an umbrella brand that sits above Parakeet rather than replacing it:

• NVIDIA's "Nemotron Speech" collection on Hugging Face, described as "Open, state-of-the-art, production-ready enterprise speech models from the NVIDIA Speech research team for ASR, TTS, Speaker Diarization and S2S," lists parakeet-unified-en-0.6b alongside nemotron-speech-streaming-en-0.6b, nemotron-3.5-asr-streaming-0.6b, the Magpie TTS model, PersonaPlex and the Granary dataset.^[18]
• NVIDIA kept publishing Parakeet-named checkpoints after the rebrand, including parakeet-ctc-0.6b-Vietnamese in January 2026 and parakeet-unified-en-0.6b in April 2026.^[14][29]
• The names cross over freely inside the model cards. multitalker-parakeet-streaming-0.6b-v1 is described as "based on the Nemotron-Speech-Streaming model"; nemotron-3.5-asr-streaming-0.6b is "the multilingual extension of nvidia/nemotron-speech-streaming-en-0.6b"; and the parakeet-unified-en-0.6b card benchmarks itself directly against nemotron-speech-streaming-en-0.6b as a sibling for different latency budgets.^[14][16][28]

Architecture does not separate the two names either. nemotron-speech-streaming-en-0.6b, released March 13, 2026, is a 600M cache-aware FastConformer-RNNT with a 24-layer encoder trained on 530,000 hours of audio, which is exactly the design used across streaming Parakeet.^[15] Its multilingual successor nemotron-3.5-asr-streaming-0.6b, released June 4, 2026, adds language-ID prompt conditioning to cover 40 language-locales from one model, split into 19 transcription-ready, 13 broad-coverage and 8 adaptation-ready languages that need fine-tuning.^[16]

The practical reading is that "Parakeet" now denotes a lineage of FastConformer CTC, RNN-T and TDT checkpoints inside a marketing family called Nemotron Speech, and that NVIDIA is not consistent about which label it attaches to a new model. Nothing NVIDIA has published retires the Parakeet name.

Scroll sideways for more →

Leaderboard figures dated July 31, 2026; the multilingual model's English WER is measured at a 1.12-second chunk in a streaming configuration, so it is not comparable with the offline entries above.^[12][16]

Benchmark standing

The Open ASR Leaderboard is re-run continuously and the results are versioned in a public dataset repository, so any claim about Parakeet's position needs a date attached. The English short-form results file dated July 31, 2026 listed 74 systems. On that snapshot:

Scroll sideways for more →

Source: Open ASR Leaderboard English short-form results, July 31, 2026.^[12]

Three things follow. First, Parakeet is no longer the accuracy leader: parakeet-tdt-0.6b-v2 sat 27th of 74 by cleaned average WER, and the best NVIDIA entry was Canary-Qwen-2.5B at 17th. The systems above it are mostly speech encoders bolted to a text LLM decoder (Qwen3, Granite, proprietary APIs), which is the pattern the leaderboard paper predicts will win on WER.^[13] Second, Parakeet's efficiency lead was untouched: among the 62 systems for which the leaderboard reports RTFx, the seven fastest were parakeet-tdt_ctc-110m (6,119), parakeet-tdt-0.6b-v3 (6,098), parakeet-tdt-0.6b-v2 (6,038), parakeet-ctc-0.6b (5,884), parakeet-rnnt-0.6b (5,407), parakeet-ctc-1.1b (5,015) and parakeet-tdt-1.1b (4,529). Third, the accuracy gap at the top is small in absolute terms (about 1 WER point between Parakeet v2 and the leader) while the throughput gap is roughly an order of magnitude, which is why Parakeet remains the default choice for batch transcription of large archives.

On the leaderboard's long-form track (Earnings-21, Earnings-22, TED-LIUM and CORAAL), in the results file dated July 10, 2026, parakeet-tdt-0.6b-v3 posted a 10.72 average, ahead of whisper-large-v3-turbo (11.01), parakeet-tdt-0.6b-v2 (11.18) and whisper-large-v3 (11.23), and behind a group of proprietary transcription APIs led by ElevenLabs Scribe v2 at 7.32. Parakeet's RTFx on that track was roughly 1,000 against 68 for Whisper large-v3.^[25]

The multilingual track compares FLEURS, Mozilla Common Voice and Multilingual LibriSpeech per language. In the files dated July 31, 2026, parakeet-tdt-0.6b-v3 beat whisper-large-v3 decisively on Common Voice in every language reported, traded results on FLEURS (better in French and Italian, worse in German, Spanish and Portuguese) and lost on MLS everywhere it was measured, while running roughly ten times faster.^[26]

Scroll sideways for more →

Word error rates in percent, Open ASR Leaderboard multilingual results, July 31, 2026.^[26]

Comparison with OpenAI Whisper

Whisper is the natural reference point, and the comparison is less about which model is better than about which constraints each was built for.

Scroll sideways for more →

Sources: Hugging Face model cards and the July 31, 2026 leaderboard file.^[6][12][24]

The differences that actually decide deployments:

• Language coverage. Whisper large-v3 is documented as covering 99 languages.^[24] Parakeet-TDT-0.6B-v3 covers 25, all European.^[6] If the input is Hindi, Arabic, Thai or Swahili, Parakeet-TDT-0.6B-v3 is not a candidate. NVIDIA's answer since June 2026 is nemotron-3.5-asr-streaming-0.6b at 40 language-locales, but even that is well short of Whisper, and 32 of the 40 transcribe out of the box, 19 marked transcription-ready and 13 broad-coverage, with the remaining 8 needing fine-tuning.^[16]
• Licensing. Both are permissive. Whisper is Apache 2.0; the CC-BY-4.0 Parakeet checkpoints require attribution. The newer Parakeet and Nemotron Speech models moved to the NVIDIA Open Model License, which permits commercial use but is a bespoke NVIDIA license rather than a standard open-source one, a distinction worth checking before shipping.^[14][15]
• Throughput. This is the largest gap. On the same leaderboard run, Parakeet transcribes roughly 13 times as much audio per unit of wall-clock time as Whisper large-v3 and about 8 times as much as the distilled whisper-large-v3-turbo.^[12] For a batch archive job that difference is the whole economic argument.
• Portability. Whisper has an enormous third-party ecosystem, including CPU-only implementations. Parakeet is designed for NVIDIA GPUs and its optimized serving path runs only on them.
• Hallucination behaviour. Encoder-decoder models such as Whisper are known to generate fluent text during silence or noise; transducer and CTC decoders do not have a text-only language-model path that can run away in the same manner. NVIDIA markets Parakeet's robustness to non-speech audio explicitly.^[1]

Relationship to Canary

Parakeet and Canary are sibling model families within NeMo that share the FastConformer encoder but diverge in decoder design and scope. Canary uses an attention-based encoder-decoder (AED) architecture, pairing a FastConformer encoder with a Transformer decoder, and is multilingual and multitask: it performs both speech recognition and speech-to-text translation between English and other supported languages. Parakeet, by contrast, uses transducer or CTC decoders and is focused purely on transcription, which tends to make it faster at inference. The two families are frequently released and benchmarked together, and in 2025 NVIDIA paired Canary-1B-v2 with Parakeet-TDT-0.6B-v3 as complementary multilingual offerings built on the shared Granary training data.

The July 2026 leaderboard snapshot shows the trade-off cleanly. canary-qwen-2.5b, which grafts a Qwen3-1.7B language model onto the Canary-1B-Flash encoder in a speech-augmented LLM design, is NVIDIA's most accurate entry at 5.06 cleaned average WER but runs at RTFx 861. parakeet-tdt-0.6b-v2 is 0.33 points worse and seven times faster.^[12][33] Where a workload is dominated by GPU-hours rather than by the last fraction of a WER point, Parakeet wins; where translation or the best possible transcript matters, Canary does.

NeMo, Riva and NIM: training versus serving

The three NVIDIA names attached to Parakeet do different jobs, and conflating them is a common source of confusion.

• NeMo is the open-source training and fine-tuning toolkit. It contains the FastConformer, RNN-T and TDT implementations, and every Parakeet model card points at it: "To train, fine-tune or play with the model you will need to install NVIDIA NeMo." A checkpoint is loaded with nemo_asr.models.ASRModel.from_pretrained(...) and transcribed with .transcribe([...]).^[6] NeMo is also where the fine-tuning recipes and the training configs used to produce the published checkpoints live.
• Riva is the optimized inference layer. It takes a NeMo checkpoint, compiles it with TensorRT, and serves it on Triton Inference Server behind a speech-specific streaming API with endpointing, punctuation and diarization attached.
• NIM microservices are prebuilt containers that package a model plus its Riva serving stack behind a stable API endpoint, so that no riva-build step is required.

The relationship between the last two changed materially. Riva release 2.24.0 records a breaking change: "Deprecated x86 data center deployments. For x86 deployments, refer to Riva ASR NIM, Riva TTS NIM, and Riva NMT NIM documentation."^[19] The current Riva SDK (2.26.0) supports embedded Jetson platforms only, specifically Jetson Thor. On x86 servers, NIM containers are therefore not one deployment option among several; they are the only supported Riva path. The same 2.24.0 release added FP8 quantization for the Parakeet-CTC English model and a Parakeet-RNNT 1.1B unified multilingual code-switch model for Jetson.^[19]

The Riva ASR NIM documentation lists Parakeet CTC (with en-US, vi-VN, es-US, zh-CN and zh-TW variants), Parakeet TDT and Parakeet RNNT alongside Nemotron ASR Streaming, Conformer CTC, Whisper Large v3 and Canary.^[20] The Parakeet weights themselves remain freely downloadable from Hugging Face and runnable through NeMo or Transformers without any of this, which is the point of the split: the models are open, the optimized serving path is NVIDIA's product.

Deployment in production

Public, documented Parakeet deployments are scarce. The most detailed is an NVIDIA case study about StudyFetch, an AI-native learning platform, titled "StudyFetch Cuts Inference Costs ~10x With NVIDIA, Expanding Access to AI Learning for Millions of Learners."^[21] It is vendor marketing about NVIDIA's own customer, every figure in it is self-reported, and none of it has been independently measured, so the claims below are reported rather than verified.

NVIDIA says StudyFetch serves more than 7 million learners and personalizes over 100 million learning interactions in real time, with its Learn Engine running on H100 and L40S GPUs. The architecture as described has Riva "running the Parakeet ASR model" producing live lecture transcripts, with NIM microservices packaging and deploying the Riva pipeline.^[21]

The headline number is narrower than it first appears, and the attribution matters:

• The claim is that "Migrating live lecture transcription onto NVIDIA Riva and Parakeet running on NVIDIA L40S GPUs in NVIDIA NIM containers delivered roughly a 10x reduction in cost on the company's largest inference workload."^[21] The reduction is attributed to the whole migration (model, serving stack and GPU choice together), not to Parakeet in isolation, and to one workload rather than to StudyFetch's inference bill overall.
• Separate distillation work on Nemotron models running on H100s is described as "on track to deliver similar substantial reductions across the rest of the inference stack", which is a forecast, not a result.^[21]
• Riva text-to-speech for the company's AI Tutor was "in development, targeted for summer rollout," so it was not in production at the time of writing.^[21]

The confound to name is not model quality. A roughly 10x cost change between a managed transcription API and self-hosted GPU inference is dominated by the difference between per-minute vendor pricing and the amortized cost of GPUs a customer already runs, not by any measured accuracy or speed difference between Parakeet and whatever it replaced (which the case study does not name). Parakeet's high RTFx makes self-hosting cheap enough to be worth doing, but the case study does not isolate that contribution.

Licensing and availability

Parakeet licensing is now split in two.

The checkpoints released between December 2023 and August 2025 (the original 0.6B and 1.1B models, the hybrid TDT-CTC models, the Japanese variant, and parakeet-tdt-0.6b-v2 and -v3) are distributed under CC-BY-4.0, which allows commercial use with attribution.^[3][4][5][6]

The checkpoints released from 2025 onward under NVIDIA's Nemotron Speech push (parakeet-unified-en-0.6b, parakeet_realtime_eou_120m-v1, multitalker-parakeet-streaming-0.6b-v1, parakeet-ctc-0.6b-Vietnamese, parakeet-rnnt-110m-da-dk and nemotron-speech-streaming-en-0.6b) carry the NVIDIA Open Model License Agreement, which permits commercial and non-commercial use but is NVIDIA's own license rather than an OSI-approved one. nemotron-3.5-asr-streaming-0.6b uses OpenMDW-1.1.^{[14][15][16][27][28][29][30]}

All of the models are available on Hugging Face and through NVIDIA's NGC catalog. They are trained and fine-tuned with the open-source NeMo toolkit, run natively in Hugging Face Transformers via the ParakeetForCTC, ParakeetForRNNT and ParakeetForTDT classes, and are packaged for production as NVIDIA NIM microservices.^[6][20][22] Interactive endpoints are available free for prototyping through build.nvidia.com.^[17]

See also

• NVIDIA Riva
• NVIDIA NIM
• NVIDIA NeMo
• NVIDIA Canary
• Whisper
• Automatic Speech Recognition Models
• Word error rate

References

1. ^NVIDIA. "Pushing the Boundaries of Speech Recognition with NVIDIA NeMo Parakeet ASR Models." NVIDIA Technical Blog, April 18, 2024. developer.nvidia.com/...h-nemo-parakeet-asr-models
2. ^Xu, H., Koluguri, N. R., Majumdar, S. "Turbocharge ASR Accuracy and Speed with NVIDIA NeMo Parakeet-TDT." NVIDIA Technical Blog, April 18, 2024. developer.nvidia.com/...h-nvidia-nemo-parakeet-tdt
3. ^NVIDIA. "nvidia/parakeet-rnnt-1.1b." Hugging Face model card. huggingface.co/...parakeet-rnnt-1.1b
4. ^NVIDIA. "nvidia/parakeet-tdt-1.1b." Hugging Face model card. huggingface.co/...parakeet-tdt-1.1b
5. ^NVIDIA. "nvidia/parakeet-tdt-0.6b-v2." Hugging Face model card. huggingface.co/...parakeet-tdt-0.6b-v2
6. ^NVIDIA. "nvidia/parakeet-tdt-0.6b-v3." Hugging Face model card. huggingface.co/...parakeet-tdt-0.6b-v3
7. ^Rekesh, D., et al. "Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition." arXiv:2305.05084, 2023. arxiv.org/...2305.05084
8. ^Xu, H., et al. "Efficient Sequence Transduction by Jointly Predicting Tokens and Durations." arXiv:2304.06795, 2023. arxiv.org/...2304.06795
9. ^NVIDIA. "NVIDIA Speech and Translation AI Models Set Records for Speed and Accuracy." NVIDIA Technical Blog, March 19, 2024. developer.nvidia.com/...rds-for-speed-and-accuracy
10. ^Hugging Face. "Open ASR Leaderboard." huggingface.co/...open_asr_leaderboard
11. ^VentureBeat. "Nvidia launches fully open source transcription AI model Parakeet-TDT-0.6B-V2 on Hugging Face." May 2025. venturebeat.com/...eet-tdt-0-6b-v2-on-hugging-face
12. ^Hugging Face. "hf-audio/open-asr-leaderboard-results", file `english_short_latest.csv`, revision dated July 31, 2026. huggingface.co/...open-asr-leaderboard-results
13. ^Srivastav, V., et al. "Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation." arXiv:2510.06961, October 8, 2025. arxiv.org/...2510.06961
14. ^NVIDIA. "nvidia/parakeet-unified-en-0.6b." Hugging Face model card. huggingface.co/...parakeet-unified-en-0.6b
15. ^NVIDIA. "nvidia/nemotron-speech-streaming-en-0.6b." Hugging Face model card. huggingface.co/...nemotron-speech-streaming-en-0.6b
16. ^NVIDIA. "nvidia/nemotron-3.5-asr-streaming-0.6b." Hugging Face model card. huggingface.co/...nemotron-3.5-asr-streaming-0.6b
17. ^NVIDIA. "NVIDIA Nemotron Speech for Developers." NVIDIA Developer. developer.nvidia.com/riva
18. ^NVIDIA. "Nemotron Speech." Hugging Face collection. huggingface.co/...nemotron-speech
19. ^NVIDIA. "Riva Release Notes." NVIDIA Riva documentation. docs.nvidia.com/...release-notes
20. ^NVIDIA. "Models" (Riva ASR NIM). NVIDIA NIM documentation. docs.nvidia.com/...models
21. ^NVIDIA. "StudyFetch Cuts Inference Costs ~10x With NVIDIA, Expanding Access to AI Learning for Millions of Learners." NVIDIA customer story. nvidia.com/...studyfetch
22. ^Hugging Face. "Parakeet." Transformers documentation. huggingface.co/...parakeet
23. ^Koluguri, N. R., et al. "Granary: Speech Recognition and Translation Dataset in 25 European Languages." arXiv:2505.13404, May 2025. arxiv.org/...2505.13404
24. ^OpenAI. "openai/whisper-large-v3." Hugging Face model card. huggingface.co/...whisper-large-v3
25. ^Hugging Face. "hf-audio/leaderboard_longform", file `longform_latest.csv`, revision dated July 10, 2026. huggingface.co/...leaderboard_longform
26. ^Hugging Face. "hf-audio/multilingual_evals", per-language result files, revision dated July 31, 2026. huggingface.co/...multilingual_evals
27. ^NVIDIA. "nvidia/parakeet_realtime_eou_120m-v1." Hugging Face model card. huggingface.co/...parakeet_realtime_eou_120m-v1
28. ^NVIDIA. "nvidia/multitalker-parakeet-streaming-0.6b-v1." Hugging Face model card. huggingface.co/...alker-parakeet-streaming-0.6b-v1
29. ^NVIDIA. "nvidia/parakeet-ctc-0.6b-Vietnamese." Hugging Face model card. huggingface.co/...parakeet-ctc-0.6b-Vietnamese
30. ^NVIDIA. "nvidia/parakeet-rnnt-110m-da-dk." Hugging Face model card. huggingface.co/...parakeet-rnnt-110m-da-dk
31. ^Andrusenko, A., Bataev, V., Grigoryan, L., Tadevosyan, N., Lavrukhin, V., Ginsburg, B. "Reducing the Offline-Streaming Gap for Unified ASR Transducer with Consistency Regularization." arXiv:2604.19079, April 21, 2026. arxiv.org/...2604.19079
32. ^NVIDIA. "NVIDIA Speech AI Models Deliver Industry-Leading Accuracy and Performance." NVIDIA Technical Blog, June 4, 2025. developer.nvidia.com/...g-accuracy-and-performance
33. ^NVIDIA. "nvidia/canary-qwen-2.5b." Hugging Face model card. huggingface.co/...canary-qwen-2.5b