Cohere has released a new speech recognition model — Transcribe

Cohere has announced a new automatic speech recognition model, Transcribe, which is now publicly available. The development is positioned as a new industry standard and is designed for use in corporate AI solutions.

Transcribe was trained from scratch with a focus on reducing error rates and ensuring practical applicability. Unlike many research models, it was originally designed as a tool for everyday tasks — from meeting transcription to speech analytics and voice assistant operations.

The model supports 14 languages, including English, French, German, Spanish, Chinese, Japanese, and Arabic. It is built on the Conformer architecture, featuring an audio processing encoder and a lightweight Transformer decoder for text generation.

According to the company, Transcribe ranks first in accuracy on the Open ASR Leaderboard on the Hugging Face platform, achieving an average error rate of just 5.42%. This outperforms solutions such as Whisper Large v3 and other speech recognition models.

In addition to high accuracy, the model demonstrates strong performance: it can quickly process audio with moderate computational resources, making it suitable for local deployment and use in business environments.

Transcribe is already available for download and can be used both locally and via API. The model also integrates with the North platform, where it will eventually become part of a broader speech analysis and processing system for corporate clients.

At Cohere, the launch of Transcribe is seen as the first step toward creating comprehensive solutions in the field of “speech intelligence,” capable of processing and analyzing voice data in real time.

Source: cohere