Cohere Transcribe

Cohere Transcribe, released on 24 March 2026 under the model name cohere-transcribe-03-2026, is Cohere's dedicated speech-recognition model: audio in, text out, two billion parameters, and nothing else. It covers 14 languages — English, German, French, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Vietnamese, Chinese, Arabic, Japanese and Korean — which puts Polish inside the supported set rather than in the long tail. The architecture is a Conformer, a speech-optimised Transformer variant: the audio waveform becomes a Mel spectrogram, a Conformer encoder holding most of the parameters processes it, and a deliberately lightweight Transformer decoder turns the result into text tokens. Cohere trained it from scratch with standard supervised cross-entropy, aiming at word error rate first. The company claims best-in-class accuracy across its 14 languages and a real-time factor up to three times faster than other dedicated ASR models of similar size. What it does not do is documented as plainly as what it does, and matters more when choosing a tool. There is no automatic language detection — the language has to be specified in advance, and the model performs best when the audio stays within it. There are no timestamps and no speaker diarisation, so subtitling and multi-speaker meeting transcripts are out of scope without a second tool. Input files are capped at 25 MB. One detail is worth naming precisely. Cohere describes this as an open-source research release under Apache 2.0, and the licence really is Apache 2.0 — but the weights repository on Hugging Face is gated: downloading requires an approved access request and an authenticated account, so the model card itself is not publicly readable. It is the reverse of the usual pattern, where the file is easy to fetch and the licence is the restrictive part. A separate Arabic-focused release, cohere-transcribe-arabic-07-2026, followed in June 2026.

▸Videos

No videos yet.

▸Reviews