IBM Granite Speech 5.0: a 2B open speech model, a faster multilingual Turbo CTC and better speaker labels

IBM published three Granite Speech updates on 21 October: one new model and revised weights for two existing ones. All three are open weights under Apache 2.0, and IBM says they run in Hugging Face Transformers. Last week IBM folded speech into a single endpoint with Granite Switch 4.1; this release is the speech models themselves.

granite-speech-5.0-2b: accuracy and translation

granite-speech-5.0-2b replaces granite-speech-4.1-2b at the same size. It transcribes English, French, German, Spanish, Portuguese and Japanese, translates between English and those languages, and also translates English to Italian and Mandarin. IBM says it was trained on 390,000 hours of public audio plus synthetic data.

New controls, per IBM:

• name the input language in the prompt to steer transcription;
• keyword lists of up to 2,000 terms to bias output toward product names, people and acronyms;
• punctuation and capitalization when you ask for it in the prompt.

IBM says the model averages 5.06% word error rate on the Open ASR Leaderboard’s English short-form sets, which it ranks second among open-weight ASR models as of 21 October. IBM lists support in Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon.

Turbo CTC: now multilingual, and faster

granite-speech-5.0-470m-turboctc is the 470M-parameter speed model IBM launched in August as English-only. The revision adds Japanese, Spanish, German, French, Italian and Portuguese, writes punctuation and capitals itself, and was trained to cope better with echoey rooms and distant microphones. IBM puts it at 19,870 RTFx on the leaderboard, up from about 12,600 at launch, “at the same accuracy level”. IBM offers a WebGPU browser demo.

4.1 2B Plus: who said what, and when

granite-speech-4.1-2b-plus keeps its name because the architecture, prompt format and API are unchanged, so IBM calls the new weights a drop-in replacement. It is aimed at meetings and calls. IBM’s table, all error rates where lower is better:

Measure	Before	Now
Long-form WER	13.53	10.22
Biased-keyword WER (20-term lists)	11.79	6.48
Average timestamp shift, all speech	210.86	98.54
Word diarization error rate	4.40	4.10

Which one to pick

IBM’s own guide: the 5.0 2B model for accuracy, translation and keyword lists; Turbo CTC for raw speed or on-device use; 4.1 2B Plus for multi-speaker recordings. Every number above is IBM’s measurement. Run a sample of your own audio, with your own vocabulary, before you switch.