Cohere Transcribe¶

Development line: project:cohere-transcribe · thread cohere-transcribe
Last event: 2026-03-27 · 1 dated since 2026-03-27 · Researched: 2026-09-05 · confidence: high

What it is¶

Cohere Transcribe is an Apache-2.0 Conformer encoder-decoder ASR model for audio-to-text workflows.

• Transcribes 14 languages with the base cohere-transcribe-03-2026 model.
• Runs locally through Transformers or vLLM, or through Cohere’s transcription API.
• Adds cohere-transcribe-arabic-07-2026 for Arabic dialects and Arabic-English audio.

Development line¶

• 2026-03-27 — Cohere publishes Cohere Transcribe model resources. On 2026-03-27, Cohere published Cohere Transcribe resources: a blog page, a Hugging Face model repository, and a demonstration space. The linked Model Vault page places the project within Cohere's model-access offering.

What changed¶

• 2026-03-27 — cohere-transcribe-03-2026 entered the product line as a 2B, 14-language open ASR model; Cohere dated the release to 2026-03-26.
• 2026-07-07 — Cohere released the separate 2B cohere-transcribe-arabic-07-2026 fine-tune for Arabic dialects, Arabic-English speech, and code-switching workloads.

How to use this¶

From 2026-03-27, evaluate Cohere Transcribe through Cohere's published model repository and demonstration space, and check the linked Cohere materials for access and deployment terms.

1. Choose cohere-transcribe-03-2026 for the 14-language base model; choose cohere-transcribe-arabic-07-2026 when the workload is predominantly Arabic or Arabic-English. — https://docs.cohere.com/docs/audio-transcription-quickstart
2. For the hosted route, obtain a Cohere API key and send a multipart request to POST /v2/audio/transcriptions with model, an ISO-639-1 language, and the audio file. — https://docs.cohere.com/reference/create-audio-transcription
3. Keep hosted input files at 25 MB or below and use FLAC, MP3, MPEG, MPGA, OGG, or WAV. — https://docs.cohere.com/docs/audio-transcription-quickstart
4. For local inference, load the base checkpoint with Transformers, pass the audio at 16 kHz and set the language explicitly; use the model card’s long-form chunking path for longer recordings. — https://huggingface.co/CohereLabs/cohere-transcribe-03-2026
5. For self-hosted production serving, follow Cohere’s vLLM deployment path and call its OpenAI-compatible audio-transcriptions endpoint. — https://huggingface.co/CohereLabs/cohere-transcribe-03-2026

Best practices¶

• Specify one expected language per request, as the base model lacks automatic language detection and produces inconsistent text on code-switched audio. — https://huggingface.co/CohereLabs/cohere-transcribe-03-2026
• Put VAD or a noise gate before the model so silence and low-level room noise do not generate text. — https://huggingface.co/CohereLabs/cohere-transcribe-03-2026
• Add a separate diarization and timestamping stage if the application requires speakers or word timing; the base model supplies neither. — https://huggingface.co/CohereLabs/cohere-transcribe-03-2026
• Use the Arabic fine-tune rather than the base model for Arabic dialects and Arabic-English speech; Cohere reports gains over the base model on both. — https://cohere.com/blog/transcribe-arabic

Superseded by this¶

• 2026-03-26 — The 2026-07-07 Arabic fine-tune supersedes the 14-language base model for Arabic dialect and Arabic-English transcription; the base model itself remains supported for its broader language set.

Still unknown¶

• No first-party Simplified Chinese announcement or Chinese-language usage documentation exists; Mandarin transcription support does not establish localized operating guidance.
• The reported leaderboard and throughput results are vendor-published measurements; we did not independently reproduce them.

Sources¶

Agent brief¶

• Subject: project:cohere-transcribe, thread cohere-transcribe, 1 dated events 2026-03-27 → 2026-03-27.
• Practical note: From 2026-03-27, practitioners should evaluate Cohere Transcribe through Cohere's published model repository and demonstration space, while checking the linked Cohere materials for access and deployment terms.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.