CohereLabs / cohere-transcribe-03-2026
cohere-transcribe-03-2026
CohereLabs cohere-transcribe-03-2026 is a 2B Conformer speech-to-text model for 14 languages. It creates accurate transcripts for meetings, calls, and audio archives.
Overview
CohereLabs built cohere-transcribe-03-2026 for automatic speech recognition. It turns spoken audio into a single, readable text transcript.
It uses a Conformer encoder with a Transformer decoder. The model converts audio into log-Mel features, then generates text tokens. This helps you turn recordings into searchable text for notes, analytics, and downstream automation.
The model supports 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese (Mandarin), Japanese, Korean, Vietnamese, and Arabic.
What you can build
- Meeting and interview transcripts for internal notes
- Call-center QA transcripts for review and coaching
- Searchable archives for podcasts, webinars, and trainings
- Voice-of-customer text corpora for topic and sentiment analysis
- Transcripts to feed into RAG pipelines and knowledge search
- Accessibility captions that start from a raw transcript
Inputs
- An audio file for transcription. WAV and MP3 work directly. If you upload M4A, FLAC, OGG, or AAC, the platform may convert it before transcription.
- A language choice using a 2-letter ISO-639-1 code. You must set the correct language. The model does not auto-detect it.
- A limit on how much text the model can generate per internal audio chunk. This helps control very long outputs on long recordings.
Outputs
The model returns a speech-to-text transcript.
- A single transcript string for the full recording.
- Text includes the recognized words in reading order.
- The output is text only. It does not include speaker labels.
Recommended settings
- For most meeting audio, keep the per-chunk text limit around 256 tokens.
- Increase the per-chunk text limit for fast speakers or dense technical content.
- Always set the language to match the recording. Don’t leave it on the wrong language.
Limitations
- No automatic language detection. Mixed-language audio can degrade quality.
- No timestamps. You can’t align words to time without extra tooling.
- No speaker diarization. Two people speaking will appear in one combined transcript.
- Silence and low-volume noise can trigger hallucinated text. Add a noise gate or voice activity detection before transcription when possible.
- Very noisy audio, far-field microphones, heavy overlap, or strong compression can reduce accuracy.
- Long recordings are typically processed in smaller chunks and then merged. Bad cuts, music, or inconsistent audio levels can create awkward joins.
- Some hosted deployments enforce an upload limit (often 25 MB). If you hit limits, compress or split the audio.
Safety & compliance
Cohere’s policies prohibit using their services for harmful or abusive content. This includes hate, sexual exploitation, and content that incites violence or self-harm.
- Get consent before recording or transcribing other people.
- Don’t submit audio that you don’t have rights to process.
- Don’t use transcripts to enable disallowed content or targeted harassment.
API quick start
Run cohere-transcribe-03-2026 with a single API call.
{
"language": "en",
"inputAudio": "https://your-cdn.com/input.mp3",
"maxNewTokens": 256
}