Cohere Transcribe

Cohere Transcribe is a family of automatic speech recognition models developed by Cohere and Cohere Labs. The first release, cohere-transcribe-03-2026, is a 2-billion-parameter audio-to-text model for 14 languages. Cohere released its weights under the Apache 2.0 license on March 26, 2026 and also made the model available through its hosted transcription API.^[1][2][3]

Cohere describes Transcribe as open source, although "open-weight" is the more precise description of what was released: model weights, configuration, inference code, and usage examples are available in a gated Hugging Face repository. A user must sign in, accept the repository conditions, and agree to share contact information before downloading its files. The license itself is Apache 2.0.^[3]

Release and variants

The March model's exact identifier is cohere-transcribe-03-2026. It accepts an audio waveform and returns transcribed text. The supported languages are English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Mandarin Chinese, Japanese, Korean, Vietnamese, and Arabic. The language must be supplied with the request because the model does not perform explicit automatic language detection.^[2][3]

On July 7, 2026, Cohere added cohere-transcribe-arabic-07-2026, a separate 2-billion-parameter fine-tune for Arabic and English. Cohere says this version emphasizes major Arabic dialects, Arabic-accented English, and Arabic-English code switching. Its model card is more cautious, stating that it still works best with a single specified language and can be inconsistent on code-switched audio. The Arabic version uses the same Conformer-based encoder-decoder design and Apache 2.0 license as the base release.^[7][8]

Both variants were live through Cohere's V2 Audio Transcriptions API when checked in July 2026. Both were also available as open weights on Hugging Face, while Cohere offered dedicated Model Vault deployments for production use.^[2][7]

Architecture and training

Transcribe is a speech-specific encoder-decoder model rather than a text large language model with audio added to it. Incoming audio is converted into a log-Mel spectrogram. A Fast-Conformer encoder extracts acoustic representations, and a smaller Transformer decoder generates text tokens through cross-attention. Cohere says more than 90 percent of the parameters are in the encoder, reducing the amount of autoregressive decoder work during inference. The model was trained from scratch with supervised cross-entropy.^[2][3][4]

The base model uses a 16,000-token multilingual byte-pair encoding tokenizer with byte fallback. Audio is automatically resampled to 16 kHz during local preprocessing, and stereo channels are averaged into one channel. Punctuation is enabled by default but can be disabled to return lower-case, unpunctuated text.^[3][4]

Cohere disclosed that the March model was trained on about 500,000 hours of curated audio-transcript pairs, augmented with synthetic data. It applied an internal cleaning pipeline, proprietary mixture balancing, background-noise augmentation from 0 to 30 dB signal-to-noise ratio, and checks for train-test audio overlap. The company did not publish a complete list of training datasets or the proportions of public, licensed, proprietary, and synthetic material. Multilingual LibriSpeech was named as one example of an open dataset used, but benchmark datasets should not be assumed to be training sources.^[4]

Inputs and output behavior

The hosted API accepts FLAC, MP3, MPEG, MPGA, OGG, and WAV files up to 25 MB. The request requires a model name, an ISO 639-1 language code, and the file. Its documented successful response contains a single text field. MP4, M4A, and WebM are not supported by Cohere's endpoint, even though some other transcription APIs accept them.^[5][6]

Cohere does not state a maximum audio duration for local inference. The official processor divides audio longer than its per-clip window into chunks and reassembles the text using chunk indexes. The model card demonstrates this with a 55-minute earnings call. The 25 MB ceiling is an API upload-size limit, not a documented local duration limit.^[3][5]

Neither the base nor Arabic model produces timestamps or speaker labels. Transcribe may recognize words in recordings with multiple speakers, and the AMI meeting corpus is included in its English evaluation, but that is different from speaker diarization. Applications that need speaker-attributed transcripts or time-aligned captions must add separate alignment and diarization components.^[2][3][8]

Local and hosted access

For offline use, Cohere recommends Transformers 5.4 or newer with PyTorch and the model's processor. The card was tested with PyTorch 2.10 and uses automatic device mapping in its examples. For production serving, Cohere recommends vLLM 0.19 with audio support. Cohere and vLLM contributors added variable-length encoder scheduling and packed decoder inputs, which Cohere says improved its serving throughput by as much as two times.^[3][4]

The model card also lists ecosystem implementations for Apple Silicon, Rust, browser WebGPU, Linux, iOS, and Android. These integrations vary in maturity and are maintained by different projects; their presence is not a guarantee of identical accuracy or performance.^[3]

Cohere has not published an official minimum VRAM, RAM, or CPU requirement. The distributed base weights use BF16 tensors and contain about 2 billion parameters, but actual memory needs include activations, audio length, batch size, runtime overhead, and any quantization. Cohere markets the model for local and edge use, and TechCrunch reported that it was intended to run on consumer-grade GPUs. That description should not be converted into a specific hardware requirement without testing.^[1][3][11]

Hosted access has two forms. Cohere offers free API experimentation subject to trial rate limits and a 25 MB upload cap. For production without those trial limits, customers can provision a dedicated Model Vault instance priced by instance-hour. Self-hosting the downloaded weights keeps infrastructure and scaling under the operator's control.^[1][2][5]

Evaluation

At release, cohere-transcribe-03-2026 recorded a mean word error rate of 5.42 on the English Open ASR Leaderboard, a March 26 snapshot that placed it just ahead of Zoom Scribe v1 at 5.47 and IBM Granite 4.0 1B Speech at 5.52. The mean covered AMI, Earnings-22, GigaSpeech, LibriSpeech test-clean and test-other, SPGISpeech, TED-LIUM, and VoxPopuli. Lower word error rate is better. The same record reported an inverse real-time factor of 524.88, meaning roughly 525 minutes of audio per minute under the leaderboard's measurement setup.^[1][3][4]

The benchmark is maintained outside Cohere. Its accompanying paper describes a reproducible evaluation of more than 60 systems across 11 datasets, with standardized text normalization, word error rate, and inverse real-time factor. Even so, a leaderboard average can hide domain-specific variation, and throughput depends on hardware, precision, batching, and runtime configuration.^[10]

Cohere also reported multilingual tests averaged over FLEURS, Common Voice 17, Multilingual LibriSpeech, and WeNet where applicable. It used character error rate for Mandarin, Japanese, and Korean and word error rate for the other languages. Its own human-preference study compared meaning preservation, named entities, hallucination avoidance, verbatim wording, and formatting. Cohere did not publish enough participant and sampling detail in the launch material to treat that study as independent. TechCrunch noted that Transcribe lost some head-to-head comparisons in Portuguese, German, and Spanish even while its reported average preference was favorable.^[4][11]

The Arabic fine-tune recorded a 25.87 average word error rate on the Open Universal Arabic ASR Leaderboard snapshot dated July 7, ahead of the listed OmniASR LLM 7B result of 28.32. Its evaluation included SADA, Common Voice, clean and noisy MASC, MGB-2, and Casablanca. These numbers apply to the Arabic variant, not the general March model.^[8]

Privacy and safety

Local inference can keep recordings off Cohere's systems, but privacy then depends on the operator's storage, logging, access controls, and downstream applications. Cohere says it does not receive customer inputs or outputs in private deployments and deployments operated through third-party cloud platforms. Model Vault is marketed as private managed inference.^[1][9]

For Cohere's SaaS platform, its enterprise data commitments say paying commercial customers can opt out of model-training use. Logged inputs and outputs are normally deleted after 30 days, with exceptions for legal or contractual requirements and suspected policy violations. Zero data retention requires approval and is limited to enterprise customers that make additional commitments. Trial-key users are governed by Cohere's general Terms of Use and Privacy Policy rather than the enterprise commitments. These are platform-wide rules, not a Transcribe-specific audio retention policy.^[9]

The most concrete model safety issue is false transcription of silence or background noise. Cohere says Transcribe can be eager to generate text from non-speech sounds and recommends placing a noise gate or voice activity detection model before it. The base model also expects monolingual input with a supplied language tag and was not explicitly trained for general code switching. These limitations can produce errors that are easy to mistake for spoken content.^[3][4]

Limitations

Transcribe performs only transcription. It does not translate speech, identify speakers, return timestamps, detect the language automatically, summarize a recording, or verify whether the generated words are faithful. Applications needing those functions require additional models and should preserve the original audio for audit where appropriate.^[2][3]

The 14-language list describes training coverage, not uniform accuracy. Performance can change with accent, dialect, microphone quality, compression, overlapping speakers, noise, specialized terminology, and recording domain. TechCrunch's review of Cohere's published comparisons found weaker relative results for several European languages, illustrating why the English leaderboard score cannot stand in for every language.^[11]

Like Whisper and other generative encoder-decoder ASR systems, Transcribe can hallucinate text during silence or noise. Critical transcripts should be checked against the recording rather than treated as a verbatim ground truth. Within the broader Voice AI field, its main advantages are permissive weights, relatively small size, multilingual support, and deployability, while its clearest gaps are missing alignment, diarization, automatic language detection, and fully disclosed training provenance.

References

1. Cohere. "Introducing Cohere Transcribe: a new state-of-the-art in open-source speech recognition." March 26, 2026. https://cohere.com/blog/transcribe ↩
2. Cohere Documentation. "Cohere Transcribe." https://docs.cohere.com/docs/transcribe ↩
3. Cohere Labs. "CohereLabs/cohere-transcribe-03-2026" model card. Hugging Face. https://huggingface.co/CohereLabs/cohere-transcribe-03-2026 ↩
4. Cohere Labs. "Introducing Cohere-transcribe: state-of-the-art speech recognition." March 26, 2026. https://huggingface.co/blog/CohereLabs/cohere-transcribe-03-2026-release ↩
5. Cohere Documentation. "Audio Transcription - quickstart." https://docs.cohere.com/v2/docs/audio-transcription-quickstart ↩
6. Cohere API Reference. "Create a transcription." https://docs.cohere.com/v2/reference/create-audio-transcription ↩
7. Cohere Documentation. "Meet Cohere Transcribe Arabic." July 7, 2026. https://docs.cohere.com/changelog/transcribe-arabic ↩
8. Cohere Labs. "CohereLabs/cohere-transcribe-arabic-07-2026" model card. Hugging Face. https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026 ↩
9. Cohere. "Enterprise Data Commitments." Updated December 5, 2025. https://cohere.com/enterprise-data-commitments ↩
10. Srivastav, Vaibhav, et al. "Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation." October 8, 2025. https://arxiv.org/abs/2510.06961 ↩
11. Ivan Mehta. "Cohere launches an open source voice model specifically for transcription." TechCrunch, March 26, 2026. https://techcrunch.com/2026/03/26/cohere-launches-an-open-source-voice-model-specifically-for-transcription/ ↩