On Thursday, the enterprise AI firm Cohere unveiled its inaugural voice model: Transcribe, an open-source automatic speech recognition model suitable for applications such as note-taking and speech evaluation.

Weighing in at only 2 billion parameters, the model is designed for use with consumer-grade GPUs for those who prefer self-hosting. It presently accommodates 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese, and Arabic.

According to Cohere, Transcribe outperforms models like Zoom Scribe v1, IBM Granite 4.0 1B, ElevenLabs Scribe v2, and Qwen3-ASR-1.7B Speech in the Hugging Face Open ASR leaderboard, registering an average word error rate (WER) of 5.42, the lowest among all models tested.

The company asserts that Transcribe achieved an average win rate of 61% against competing models when human judges evaluated its transcriptions based on accuracy, coherence, and usability. Nevertheless, the model lagged behind its competitors in transcribing Portuguese, German, and Spanish.

Cohere notes that Transcribe can handle 525 minutes of audio in just one minute, which is impressive for its model category.

The firm intends to incorporate Transcribe into its enterprise agent orchestration platform, North, and is offering the model for free via its API. Additionally, it will be accessible on Model Vault, Cohere’s managed inference platform.

The popularity of speech recognition models is on the rise as the need for note-taking and dictation applications like Granola and Wispr Flow increases.

Earlier this year, Cohere reportedly informed investors that it was on track for an annual recurring revenue of $240 million by 2025, with CEO Aidan Gomez stating that the startup might go public “soon”.