IBM Granite Speech gains a new 2B model that transcribes English, French, German, Spanish, Portuguese and Japanese, translates to and from English in those same languages, and adds English into Italian and Mandarin. The model, Granite Speech 5.0 2B, carries about 2.3 billion parameters and ships under the Apache 2.0 license, with support in Transformers (version 4.52.1 or later), vLLM, llama.cpp and mlx-audio for Apple Silicon. IBM advises feeding it audio in chunks of up to about two minutes and splitting longer recordings, for example with voice activity detection, before decoding.
It succeeds granite-speech-4.1-2b without growing in size. IBM reports an average word error rate of 5.06% across the Open ASR Leaderboard's public English short-form test sets, which it says ranks second among top open-weight models as of October 21. IBM trained it on 390,000 hours of public audio plus synthetic data for Japanese recognition, keyword biasing and translation, starting from an intermediate checkpoint of its granite-4.0-1b-base language model.
IBM says the new model is more accurate than 4.1 on English short-form benchmarks and handles noise, reverberation and distant-microphone speech better. A prompt can now name the input language to steer multilingual transcription. Keyword lists now run to 2,000 terms in every supported language, which lets a team bias output toward product names, people and acronyms. Punctuation and capitalization are available on request; capitalization F1 for the 5.0 model lands between 90.2 and 99.5 on IBM's test sets.
The release covers three checkpoints in IBM's Granite Speech collection. The 2B is the only new model; the other two are updates:
| Model | Status | Aimed at |
|---|---|---|
| granite-speech-5.0-2b | New | Accurate multilingual transcription and translation |
| granite-speech-5.0-470m-turboctc | Updated | Fast and on-device transcription |
| granite-speech-4.1-2b-plus | Updated | Meetings and multi-speaker calls |
The 470-million-parameter Turbo CTC encoder, English-only when it launched in August, now handles English, Japanese, Spanish, German, French, Italian and Portuguese, and produces punctuated, capitalized text natively. By compiling a single Conformer block and reusing its graph across the stack, IBM raised its leaderboard throughput from roughly 12,600 to 19,870 RTFx, about 1.6 times faster at the same accuracy.
The updated granite-speech-4.1-2b-plus targets long, multi-speaker recordings. Word error rate on long-form test sets falls from 13.53 to 10.22. With 20-term keyword lists, errors on biased keywords drop from 11.79 to 6.48, and the average timestamp shift is cut by more than half. Because the architecture, prompt format and API are unchanged, IBM calls it a drop-in replacement for the current 4.1 weights.
