What is Cohere Transcribe?
Cohere Transcribe is Cohere's first audio model: a 2B-parameter Conformer encoder with a small text decoder, trained from scratch on about 500,000 hours of audio. You send a recording and a language code, and it returns punctuated plain text. The weights are free to download under Apache 2.0, and Cohere's API gives a free, rate-limited trial.
It is Cohere's current speech model, with an Arabic fine-tune alongside it. Cohere reports a lower error rate than Whisper Large v3 (5.42% against 7.44%) and higher throughput, but unlike ElevenLabs Scribe v2 or Mistral's Voxtral it gives no timestamps, speaker labels or language detection.
How is Cohere Transcribe Arabic different?
It is a free fine-tune of Cohere Transcribe for Arabic dialects and mixed Arabic-English speech, with the same size, licence and limits.
- Languages
- Arabic and English, including regional dialects and switching between the two mid-sentence.
- Accuracy
- Cohere reports a 25.87% average word error rate on the Open Universal Arabic ASR leaderboard, against 30.67% for the base model.
- Access
- Its own model ID on Cohere's API and its own Apache 2.0 weights on Hugging Face.
- Same limits
- No timestamps or speaker labels, and production hosting through Cohere runs on Model Vault.
Who is Cohere Transcribe best for?
Cohere Transcribe is best for developers and companies that want accurate, free speech-to-text on their own hardware, for dictation, audio search and single-language recordings in 14 languages.
Self-hosted transcription
Free Apache 2.0 weights and a 2B size let you run it on your own GPUs or at the edge, so sensitive audio never leaves your systems.
Dictation and voice notes
Its model card lists builds for Apple Silicon, the browser and dictation apps on iOS, Android and Linux, and it punctuates by default.
Searchable audio archives
Batch transcripts of calls, lectures and training recordings can feed search indexes and RAG pipelines, with long files split and stitched back automatically.
Single-language business audio
It covers 14 languages across Europe, Asia and the Middle East, as long as each recording sticks to one language you name up front.
What's included with Cohere Transcribe?
Free Apache 2.0 weights, a free trial API, automatic handling of long recordings, a punctuation switch, vLLM serving, managed Model Vault hosting and an Arabic variant.
- Open weights
- Apache 2.0 weights on Hugging Face, where you log in and agree to share contact details before downloading.
- Transcription API
- Cohere's audio transcriptions endpoint takes FLAC, MP3, OGG or WAV files up to 25MB, free on a trial key.
- Long recordings
- The processor splits audio longer than one chunk and reassembles the text; Cohere's own example transcribes a 55-minute earnings call.
- Punctuation switch
- Punctuated, cased text by default, or lower-case text without punctuation when you turn it off.
- vLLM serving
- Cohere added support to vLLM for production serving, which it says gives up to twice the throughput.
- Model Vault hosting
- A dedicated, single-tenant instance run by Cohere, with no rate limits, in Medium or Large tiers.
- Arabic variant
- Cohere Transcribe Arabic is a free fine-tune for Arabic dialects and Arabic-English code-switching.
How much does Cohere Transcribe cost?
Cohere Transcribe is free to download and run under Apache 2.0 and free to test on a rate-limited trial key; managed production hosting on Model Vault starts at $3.75 per instance per hour.
- Apache 2.0 licence, commercial use allowed
- Runs on your own GPUs or Apple Silicon
- Hugging Face login and contact details needed
- 5 requests a minute, up to 1,000 API calls a month
- Files up to 25MB
- Not for production or commercial use
- $2,500 per instance on a monthly plan
- Dedicated instance with no rate limits
- Autoscaling on Flex plans
- $4,750 per instance on a monthly plan
- Higher-throughput tier
- Pause an instance to stop paying for it
Cohere lists no pay-per-minute price for Transcribe: production use through Cohere runs on a Model Vault instance, and speech models there may need a waitlist. Annual Model Vault plans cost $25,000 (Medium) or $47,500 (Large) per instance, and Encrypted Vaults with confidential computing are priced separately. The model works on audio in 30-second chunks. Full pricing on cohere.com. We checked these prices on .
How do readers rate Cohere Transcribe?
No one has rated Cohere Transcribe yet. Be the first.
Sign in to rate. One rating per account.
Cohere Transcribe pros and cons
It is accurate, free and fully self-hostable, but it returns plain text only, with no timestamps, speaker labels or language detection.
Pros
- Cohere reports a 5.42% average word error rate on the Open ASR Leaderboard, and several Reddit users who moved from Whisper or Parakeet for dictation say it is both faster and more accurate on their own recordings.
- Apache 2.0 weights allow commercial use and fully local processing, which Hacker News commenters single out for meetings that cannot be sent to an outside service.
- Clean punctuation out of the box: testers at Whisperian and Superwhisper users both say it punctuates better than the models they used before.
- Free to try two ways: download the weights, or call Cohere's API on a trial key with no setup beyond an account.
Cons
- No timestamps and no speaker labels, the most common complaint on Hacker News and Reddit, which rules it out for subtitles and multi-speaker meeting notes without extra tools.
- You must name the language of every recording: there is no automatic detection, and Cohere says mixed-language audio gives inconsistent results.
- Heavier and slower than NVIDIA's Parakeet on phones and laptops, and results vary by task; one Hacker News tester's British postcode set put it behind Soniox, ElevenLabs and AssemblyAI.
- No pay-as-you-go production API: past the free trial, Cohere hosting means a dedicated Model Vault instance from $3.75 an hour, and speech models may need a waitlist.
Is Cohere Transcribe open source?
Yes. Cohere Transcribe's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Cohere Transcribe alternatives?
Cohere Transcribe is an open source alternative to Scribe v2.
Scribe v2FreeChoose it when you need more than 90 languages with speaker diarization and word-level timestamps from a hosted API; ElevenLabs charges $0.22 an hour of audio.
Voxtral Transcribe 2FreemiumChoose it when you need speaker labels, word-level timestamps or custom vocabulary; Mistral's batch API costs $0.003 a minute, and its Realtime model is also open weights under Apache 2.0.
Gemini 3.5 TranscribeFreeChoose it when recordings mix languages; it detects 85+ languages on its own, labels speakers and has a free tier, then costs about $0.005 a minute.
Questions about Cohere Transcribe
What is Cohere Transcribe?
Cohere Transcribe is Cohere's open-weights 2B speech recognition model that turns recorded audio in 14 languages into text, for developers and companies that want transcription running on their own hardware.
Who is Cohere Transcribe for?
Cohere Transcribe is best for developers and companies that want accurate, free speech-to-text on their own hardware, for dictation, audio search and single-language recordings in 14 languages.
Is Cohere Transcribe free?
Yes. Cohere Transcribe has a free Open weights plan: apache 2.0 licence, commercial use allowed.
How much does Cohere Transcribe cost?
Cohere Transcribe's paid plans are Model Vault Medium at $3.75 per instance per hour and Model Vault Large at $7.50 per instance per hour.
Is Cohere Transcribe open source?
Yes. Cohere Transcribe's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Cohere Transcribe alternatives?
The closest alternatives on Recaply are Scribe v2, Voxtral Transcribe 2 and Gemini 3.5 Transcribe.
Where can I try Cohere Transcribe?
On its official website, cohere.com.
Can I use Cohere Transcribe for commercial work?
Yes, when you run the open weights yourself: Cohere released the model under the Apache 2.0 license, which allows commercial use. Cohere's free trial API key is for testing only, so production use through Cohere needs a Model Vault instance, from $3.75 an hour.
Which languages does Cohere Transcribe support?
Cohere Transcribe was trained on 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Mandarin Chinese, Japanese, Korean, Vietnamese and Arabic. You set the language for each file, since it does not detect it, and Cohere says accuracy is strongest in English.
Does Cohere Transcribe do timestamps or speaker diarization?
No. Cohere lists both as limitations of the model, and the Arabic version has the same gap. For subtitles or meeting notes that need who said what and when, pair it with a separate alignment or diarization tool, or pick a model such as Voxtral Transcribe 2 or Scribe v2.
Is Cohere Transcribe better than Whisper?
On Cohere's published numbers it is more accurate: a 5.42% average word error rate on the Open ASR Leaderboard against 7.44% for Whisper Large v3, with higher throughput. It covers only 14 languages and has no timestamps, so check both before replacing Whisper in an existing pipeline.
Why does Cohere Transcribe write text when nobody is speaking?
Like most encoder-decoder speech models, it tries to transcribe any sound, including background noise and near silence. Cohere recommends putting a noise gate or voice activity detection step in front of it so quiet stretches do not turn into invented words.
Our verdict
Cohere Transcribe is one of the most accurate speech-to-text models you can run yourself for free, as long as plain text is all you need.
Cohere's published error rate beats Whisper Large v3 and Scribe v2, the weights are Apache 2.0, and users who moved from Whisper or Parakeet mostly praise its accuracy and punctuation.
- Skip it for
- Subtitles, multi-speaker meeting notes or mixed-language calls, since it has no timestamps, speaker labels or language detection.
- Check first
- Run it on your own recordings with voice activity detection in front, and ask Cohere about Model Vault capacity before planning hosted production.
MAI-Transcribe-2-Streaming
ZONOS2
Qwen3-TTS
Fish Audio S2 Pro
MisoTTS
Eleven v4