What Is Cohere Transcribe
Cohere Transcribe is the speech recognition model Cohere released on March 26, 2026 as cohere-transcribe-03-2026, its first model built for transcription. It was trained from scratch on about half a million hours of curated audio and transcripts, and it has roughly 2 billion parameters, most of them in a Fast-Conformer encoder paired with a small decoder. That design keeps it quick: it reads audio in one pass and spends little time generating text.
The weights are open under Apache 2.0, which allows commercial use, and download from the Hugging Face model card after you accept a contact-sharing form. The model covers 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese, and Arabic. In July 2026 Cohere added an Arabic fine-tune that also handles Arabic and English mixed in one recording.
The launch drew attention because an enterprise AI company released a competitive speech model with open weights. Within weeks, community ports brought it to Apple Silicon, browsers, and phones, and dictation apps started shipping it as a local model.
Accuracy and Benchmarks
At launch, Cohere Transcribe took first place on the Hugging Face Open ASR Leaderboard with a 5.42% average word error rate, ahead of Zoom Scribe v1, IBM Granite, Qwen3-ASR, ElevenLabs Scribe v2, and Whisper Large v3 at 7.44%. It also processed audio about 525 times faster than real time on the leaderboard's GPUs.
Several models now rank ahead of it: Zoom Scribe v2 Pro leads at 4.38%, Qwen3-ASR-1.7B is the best open model at 4.95%, and NVIDIA Canary-Qwen 2.5B scores 5.23%. The leaderboard has also changed its test sets and text normalization several times since March, so compare models within one snapshot, not across dates. In the snapshot from September 25, 2026, Cohere Transcribe ranks 20th of 69 English models at 5.40%. That puts it next to Parakeet TDT 0.6B v2 (5.48%) and well ahead of Whisper Large v3 (6.50%).
The gap widens outside English. On the leaderboard's multilingual data, Cohere Transcribe beats Parakeet TDT 0.6B v3 in all six European languages scored there, with Dutch at 4.17% against 7.53%, and beats Whisper Large v3 in German, French, Italian, Spanish, and Dutch. These scores use the full-precision weights; quantized builds like the ones dictation apps ship trade a little accuracy for size.
In Cohere's own human evaluation, listeners preferred Cohere Transcribe's English transcripts over those of competing models 61% of the time on average: 51% against ElevenLabs Scribe v2, 64% against Whisper Large v3, and 78% against IBM Granite 4.0 1B. These are Cohere's numbers, not a third-party result.
Cohere Transcribe vs Whisper vs Parakeet: Which to Use
Local dictation in European languages
Cohere Transcribe. On the leaderboard's multilingual data it beats Parakeet TDT 0.6B v3 in German, French, Italian, Spanish, Portuguese, and Dutch, and Whisper Large v3 in all but Portuguese.
Local English dictation
Close call. Cohere Transcribe edges Parakeet TDT 0.6B v2 on the English leaderboard (5.40% vs 5.48%), while Parakeet runs more than six times faster on the leaderboard's GPUs.
Live text while you speak
Spokenly's streaming local models, Parakeet EOU and Nemotron, show words as you speak. Cohere Transcribe only processes finished audio.
Languages beyond the 14
Whisper Large v3 Turbo, which covers about 99 languages and runs on Apple Silicon and Intel Macs, Windows, and Linux.
Subtitles, timestamps, or speaker labels
A Whisper-based pipeline or a cloud model that returns timing and speakers. Cohere Transcribe outputs plain text only.
Windows or Linux
Parakeet or Whisper locally in Spokenly, or Cohere through your own API key. Handy bundles a local Cohere build on Windows and Linux, and Superwhisper Pro does on x64 Windows.
The Parakeet vs Whisper comparison covers the two older local engines in depth, and the Whisper model size guide maps each Whisper size to the hardware it needs.
Cohere Transcribe Limitations
No timestamps or speaker labels
The model returns plain text. Cohere has said timestamps are planned for a future version; community fine-tunes add them experimentally.
No streaming
It transcribes finished audio (transformers splits long files into 35-second chunks by default), so live captions need a different model.
No official language detection
You normally pass the language with each request; some local builds use the model's unknown-language prompt as a workaround. A wrong language tag produces fluent text in the wrong language instead of an error, and mixed-language speech is handled inconsistently.
No custom vocabulary or prompts
You cannot bias it toward names or jargon, so rare terms need manual fixes or a word replacement step.
Hallucinations on silence and noise
Like Whisper, it can invent text for non-speech audio. Apps that ship it usually add voice activity detection in front.
How to Run Cohere Transcribe
On your own hardware
Hugging Face transformers 5.4 or later loads the model for offline inference, and vLLM 0.19 serves it online on a GPU. Community ports cover the rest: mlx-audio for MLX on Apple Silicon, FluidAudio for CoreML, ONNX builds for sherpa-onnx and the browser, and GGUF builds for the transcribe.cpp runtime (stock whisper.cpp does not load it). Memory depends on the runtime: one MLX run peaked at 4.5 GB at full precision, and 4-bit and 8-bit builds use roughly 2.5 to 3.3 GB. The transcribe.cpp port reports about 75 times real time on an M4 Max GPU.
Through Cohere's API
The transcription endpoint is POST https://api.cohere.com/v2/audio/transcriptions. It requires the model and language fields and accepts FLAC, MP3, OGG, or WAV files up to 25 MB. Trial keys are free but capped at 5 requests per minute and 1,000 calls per month, and Cohere has not published a per-minute production price. Teams that need dedicated capacity can host it in Cohere's Model Vault from $3.75 per hour.
Cohere Transcribe in Spokenly
Spokenly added an on-device Cohere model for dictation in June 2026 and the lighter Cohere Transcribe Q4 in August. Both run locally, so your audio never leaves the device. Three routes are available:
Cohere Transcribe Q4 · Local, free, Mac and iPhone
About a 1 GB download that holds roughly 1.7 GB of memory while loaded. 14 languages, with automatic detection through the model's own unknown-language prompt, or pick the language yourself. Needs an Apple Silicon Mac with macOS 15 or later, or an iPhone with iOS 18 or later. On macOS 15 it cannot use the GPU and runs more slowly; macOS 26 runs it at full speed.
Cohere Transcribe (INT8) · Local, experimental, Mac
A 4.7 GB CoreML build that uses around 10 GB of RAM while it runs, far more than community 8-bit builds on other runtimes. It needs a fixed language. Meant for Macs with memory to spare, not for iPhone.
Cohere API · Bring your own key, every platform
Paste a Cohere API key and Spokenly sends recordings to Cohere's transcription endpoint, on Mac, iPhone, Windows, and Linux. It needs a fixed language. Trial keys allow 1,000 calls a month, about 33 dictations a day.
Local models are free with no word caps. With Local Only Mode on, Spokenly blocks outbound network traffic while you dictate. Word replacements make up for the missing custom vocabulary: map a misheard name to its correct spelling once, and Spokenly fixes it in every transcript. On Windows and Linux, local dictation uses Parakeet and Whisper, and Cohere is available through your own API key.
Other Apps That Support Cohere Transcribe
Handy, which is free and open source, added Cohere Transcribe in April 2026. Superwhisper, which announced a partnership with Cohere in late July, offers the model as a Pro local option on macOS, Windows, and iOS, and added automatic language detection and its own vocabulary layer in August. VoiceInk ships an experimental version, OpenWhispr added Cohere Transcribe in September, and several open-source Linux dictation tools support it too. MacWhisper does not offer it. For a wider view of the category, see the best dictation software roundup.
Cohere Transcribe FAQ
What is Cohere Transcribe?
Cohere Transcribe is an open-weights speech recognition model that Cohere released on March 26, 2026 as cohere-transcribe-03-2026. It has about 2 billion parameters, transcribes 14 languages, and ships under the Apache 2.0 license, so you can run it on your own hardware or call it through Cohere's API.
Is Cohere Transcribe free and open source?
The weights are free under Apache 2.0, which allows commercial use; Hugging Face asks you to accept a contact-sharing form before downloading. Running it locally costs nothing beyond your hardware. Cohere's hosted API has a free trial tier limited to 5 requests per minute and 1,000 calls per month.
Which languages does Cohere Transcribe support?
Fourteen: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese, and Arabic. The model does not detect the language officially, so you pass it with each request, and it handles mid-sentence language switches inconsistently. A separate Arabic fine-tune released in July 2026 handles Arabic and English mixing.
Is Cohere Transcribe better than Whisper?
On the Hugging Face Open ASR Leaderboard its full-precision weights beat Whisper Large v3 on English, 5.40% against 6.50% average word error rate as of September 25, 2026, and also score better in German, French, Italian, Spanish, and Dutch. Whisper still covers far more languages, returns timestamps, and is easier to find in existing tools.
Does Cohere Transcribe support timestamps or speaker labels?
No. The March 2026 model returns plain text without timestamps or speaker diarization, and it does not stream partial results. Cohere has said timestamps are planned for a future version. For subtitles or speaker-labeled transcripts, use a Whisper-based pipeline or a cloud model that returns them.
How do I run Cohere Transcribe on a Mac?
One route is Spokenly: download the Cohere Transcribe Q4 local model (about 1 GB) on an Apple Silicon Mac with macOS 15 or later and dictate with it system-wide. Handy, Superwhisper, and VoiceInk also ship local builds, and developers can run the Hugging Face weights through MLX, CoreML, ONNX, or GGUF ports, or with transformers and vLLM on a GPU.
Does Cohere have a transcription API, and what does it cost?
Yes. Send a POST request to https://api.cohere.com/v2/audio/transcriptions with the model and language fields; it accepts FLAC, MP3, OGG, or WAV files up to 25 MB. Trial keys are free with rate limits, and Cohere has not published a per-minute production price. Dedicated hosting through Cohere's Model Vault starts at $3.75 per hour.
Does Spokenly support Cohere Transcribe?
Yes. On Mac and iPhone, Spokenly offers Cohere Transcribe Q4 as a free local model with automatic language detection, and Macs with plenty of memory can also run an experimental full INT8 model. On every platform, including Windows and Linux, a Cohere provider lets you use your own API key.