High-Performance Speech Transcription Model Transcribe

Cohere has released Transcribe. This open-source ASR model is a 2B-scale conformer-based encoder-decoder architecture that converts speech waveforms into a log-Mel spectrogram before generating text.

The core goal was to lower WER (Word Error Rate) in real-world usage environments. The model was trained from scratch, with operational suitability, inference efficiency, and deployment flexibility all considered together so that it wouldn't remain merely a research output.

It supports a total of 14 languages.

• European: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish
• AIPAC: Chinese(Mandarin), Japanese, Korean, Vietnamese
• MENA: Arabic

In terms of performance, it ranked #1 on Hugging Face's Open ASR Leaderboard with an average WER of 5.42%. Cohere emphasized that this result holds up even under realistic conditions such as conference-room audio, multi-speaker environments, and diverse accents.

Comparison targets included both open and closed ASR models such as Whisper Large v3, ElevenLabs Scribe v2, and Qwen3-ASR-1.7B. Cohere also explained that in human evaluations, it showed generally superior transcript quality in English and some of the other supported languages.

On the speed front, it maintains high throughput (RTFx) among 1B+ parameter-class models, putting the balance of accuracy and throughput front and center. In other words, its positioning targets both low WER and high processing speed at the same time.

There are also multiple deployment paths.

• Downloadable from Hugging Face
• Free experimentation via the Cohere API (rate limits apply)
• For production use, Model Vault supports low-latency, private cloud inference

Cohere stated that going forward, through integration with North, it plans to expand Transcribe beyond a simple transcription model into the foundation for enterprise speech intelligence.