In short
- IBM Granite released three speech models under Apache 2.0: a 2B, a 470M and a meetings version.
- The accuracy and speed figures come from IBM, not an independent test.
The short version: the IBM Granite team published three speech-to-text models on the Hugging Face blog. All are under Apache 2.0, so you can use them commercially and run them yourself.
What changed
- granite-speech-5.0-2b (about 2B parameters): transcription in English, French, German, Spanish, Portuguese and Japanese, translation to and from English, and keyword lists of up to 2,000 terms. IBM reports an average WER of 5.06% and 2nd place among open models on the Open ASR Leaderboard as of 21 October 2026.
- granite-speech-5.0-470m-turboctc (470M): built for speed and on-device use; IBM reports throughput rising from about 12,600 to 19,870 RTFx, roughly 1.6x.
- granite-speech-4.1-2b-plus: for multi-speaker meetings; IBM says long-form error fell 24% and errors on keywords fell 45%.
How to choose
IBM suggests: accuracy, translation and keywords call for the 2B; raw speed or local devices call for the 470M; multi-speaker meetings call for 4.1-2b-plus. The models run in Transformers, vLLM, llama.cpp and, on Macs, mlx-audio.
How to read this
All figures are IBM’s own reported results, not independent verification. Polish is not on the language list, so test your own recordings before trusting the output. More on audio and video in AI is in the image, video and voice hub.