cohere-transcribe Series
Notes
- License: Apache 2.0, available for commercial use and redistribution
- Base model sourced from
CohereLabs/cohere-transcribe-03-2026, a 2B-parameter Conformer encoder + lightweight Transformer decoder. Trained from scratch on 500k hours of high-quality audio-text pairs. It delivers up to 3× higher inference throughput compared to other ASR models of similar parameter size.- Input preprocessing: Audio is automatically resampled to 16kHz; multi-channel stereo inputs are averaged to mono and converted to log-Mel spectrograms. Long audio is automatically chunked and reassembled, supporting hour-long transcription tasks.
- Output features: Punctuation is enabled by default and can be disabled via parameter.
- Limitations: No built-in timestamps, speaker diarization, or automatic language detection. The target language must be manually specified before inference. Performance is unstable on code-switched audio. Background noise may be hallucinated as transcribed text; VAD (Voice Activity Detection) is recommended as a preprocessing step.
- Model performance: Achieves an overall average WER of 5.42% on the official ASR leaderboard. It delivers strong accuracy for meeting recordings and long-form audio, outperforming mainstream open-source ASR models such as Whisper Large v3.
- Ecosystem support: Natively compatible with Transformers, vLLM, MLX, Rust and WebGPU. Deployable on web browsers, Apple Silicon, servers and other platforms.
Full supported language & locale mapping: