IBM releases a free AI that transcribes your meetings in French without an online service

Can an AI transcribe a meeting without the recording ever leaving the company? IBM wants to say yes. On October 20, the American group published Granite Speech 5.0 on the Hugging Face platform, a speech recognition model with two billion parameters. It is released under the Apache 2.0 license: you can download it, modify it and use it for free, including for commercial purposes.

The model transcribes English, French, German, Spanish, Portuguese and Japanese. It translates these languages into English, and English into each of them, as well as into Italian and Mandarin. IBM presented it on October 21 on the Hugging Face blog, along with two other updated models.

Three models for three uses

• Granite Speech 5.0 2B, the newcomer, handles transcription and translation, with punctuation and capitalization on request. It accepts a list of up to 2,000 keywords to recognize proper names, products or in-house acronyms.
• Granite Speech 5.0 470M Turbo CTC, smaller and built for speed, now transcribes seven languages, including French and Italian; it previously handled only English. IBM mentions a demo that runs directly in the browser.
• Granite Speech 4.1 2B Plus, updated, targets meetings, support calls and multi-speaker interviews, attributing each utterance to its speaker.

These models work with common tools (Transformers, vLLM, llama.cpp, and mlx-audio for Macs with Apple silicon), so they can run on your own machines. IBM does not specify the memory required. The main model was trained for 40 days on eight NVIDIA H100 cards, using 390,000 hours of public audio.

Good scores… measured in English

IBM highlights an average word error rate of 5.06% in English. According to IBM, which cites official results as of October 21, this score places the model second among open models on the Open ASR Leaderboard, a public speech recognition ranking hosted by Hugging Face. The ranking can be viewed by anyone, but this score covers only its English tests.

For French, the picture is blurrier. The model card shows multilingual error rates only as charts, with no figures in the text. The only readable French figures concern punctuation and capitalization, measured on the public CommonVoice corpus: the new model does slightly worse there than its predecessor (11.43% punctuation errors versus 11.00%, and 97.20 versus 97.25 for capitalization).

The model card points out other limits. Audio must be split into segments of two minutes at most. Compressed versions (known as quantized) lose accuracy beyond about one minute; IBM therefore recommends the Q8_0 format with llama.cpp. And the model can, in some cases, produce inaccurate, biased or undesirable transcriptions. IBM advises pairing it with Granite Guardian, its risk detection model.

Should you try it?

A company that wants to transcribe meetings or calls without entrusting the audio to an online service now has one more option, free and modifiable. The right test: a few hours of your own recordings, with your accents, your background noise and your jargon, compared with the tool you already use. Voice is also attracting European vendors: Mistral promised on October 21 voice agents able to answer the phone. It remains to be seen what independent evaluations will show in French.