Cohere Transcribe
Cohere's first open speech model: a 2B transcription specialist for 14 languages, released under Apache 2.0.
In short. Cohere Transcribe turns recorded speech into text in 14 languages, including English, Chinese, Japanese, Korean and Arabic. It does one job, transcription, and is built to be fast and accurate rather than to chat. You tell it which language to expect, and it does not add timestamps or label speakers.
Summary
Cohere released Cohere Transcribe, its first voice model, on 26 March 2026 under Apache 2.0. It has about 2.07 billion parameters and is a dedicated audio-in, text-out speech recognizer trained from scratch, not a language model with an audio add-on. A Fast-Conformer encoder holds more than 90% of the parameters and a lightweight Transformer decoder writes the text, so little compute is spent on token-by-token generation. It was trained on 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Mandarin Chinese, Japanese, Korean, Vietnamese and Arabic. Audio is resampled to 16 kHz, stereo is averaged to one channel, and punctuation can be switched off with a flag.
Cohere reports an average word error rate of 5.42% on the English Open ASR Leaderboard as of 26 March 2026, which was first place at the time. Its table lists Zoom Scribe v1 at 5.47%, IBM Granite 4.0 1B Speech at 5.52%, NVIDIA Canary Qwen 2.5B at 5.63%, Qwen3-ASR-1.7B at 5.76%, ElevenLabs Scribe v2 at 5.83% and Whisper Large v3 at 7.44%. Cohere also says offline throughput is up to three times that of similarly sized dedicated recognizers, and reports strong results in human-preference tests of transcript quality. A newer model on this site, Granite Speech 5.0 2B, has since reported a lower English average. All of these figures come from the developers and the leaderboard moves often.
The Hugging Face card gives three limits. The model works best when you name one of the 14 languages in advance, has no automatic language detection and does not handle code-switched audio consistently. It does not produce timestamps or speaker labels. And like most encoder-decoder recognizers it is eager to transcribe, so quiet noise can turn into invented text unless you put a noise gate or voice-activity detector in front of it.
It runs in Hugging Face Transformers from version 5.4.0, in vLLM for serving through an OpenAI-style transcription endpoint, and in mlx-audio on Apple silicon, and the card lists community ports such as a Rust version, a browser demo and GGUF builds. Long recordings are split into chunks and stitched back together automatically. The weights sit behind a Hugging Face access prompt that asks you to share contact details, and Cohere also offers the model through its API and a managed service.
Good for
- Accurate transcription of calls, meetings and interviews in any of the 14 supported languages
- Self-hosted speech recognition with an Apache 2.0 license and no per-minute fees
- Batch transcription of long recordings, since the model is built for high offline throughput
- Products in Chinese, Japanese, Korean or Arabic, which many European-focused recognizers do not cover
Watch out for
- No automatic language detection, so you must pass the language, and code-switching is handled inconsistently
- No timestamps and no speaker labels, so use another tool if you need subtitles or who-said-what
- Quiet noise can be transcribed as words unless you add voice-activity detection first
- The first-place ranking dates from March 2026 and the leaderboard has moved since, and all figures are Cohere's own
- Hugging Face asks you to share contact details before it releases the files
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Cohere Transcribe. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Cohere Transcribe 03-2026 BF16 | 2.07B | 4.13 GB | 4.75 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Parakeet TDT 0.6B v3 vs Cohere Transcribe: which should you use?
- Best open-source speech models for transcription and text-to-speech
- Best open-source speech recognition (speech-to-text) models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
Frequently asked questions
- What is Cohere Transcribe?
- Cohere Transcribe turns recorded speech into text in 14 languages, including English, Chinese, Japanese, Korean and Arabic. It does one job, transcription, and is built to be fast and accurate rather than to chat. You tell it which language to expect, and it does not add timestamps or label speakers.
- Can I use Cohere Transcribe commercially?
- Yes. Cohere Transcribe is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Cohere Transcribe open source?
- Cohere Transcribe is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Cohere Transcribe on my own computer?
- Yes. The smallest variant, Cohere Transcribe 03-2026, needs about 4.75 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Cohere Transcribe released?
- Cohere Transcribe was released by Cohere on March 26, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Cohere: Introducing Cohere Transcribe (26 Mar 2026) cohere.com
- Hugging Face blog: Introducing Cohere-transcribe (26 Mar 2026) huggingface.co
- cohere-transcribe-03-2026 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Granite Speech 5.0 IBM | 470M – 2.31B | – | Apache 2.0 (Permissive) | 1.09 GB Laptop or phone | |
| Parakeet TDT 0.6B v3 NVIDIA | 627M | – | CC BY 4.0 (Permissive) | 2.89 GB Laptop or phone | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Whisper large-v3-turbo OpenAI | 810M – 1.54B | – | MIT (Permissive) | 6 GB Laptop or phone |