IBM Unveils Granite Speech 5.0 Models for Multilingual Transcription and Translation

IBM has released new additions to its Granite Speech collection: a new speech model, granite-speech-5.0-2b, and updated versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The models target enterprise use cases including high-accuracy transcription, translation, and multi-speaker recording analysis.

granite-speech-5.0-2b succeeds granite-speech-4.1-2b at the same parameter count. It transcribes English, French, German, Spanish, Portuguese, and Japanese, and translates to and from English in those languages, plus English-to-Italian and English-to-Mandarin. The model was trained on 390,000 hours of public audio and synthetic data, with improvements in noise and reverberation robustness, language-ID prompting, long-form translation, and keyword biasing of up to 2,000 terms. It is released under Apache 2.0 and supports Transformers, vLLM, llama.cpp, and mlx-audio.

The updated granite-speech-5.0-470m-turboctc expands from English-only to seven languages: English, Japanese, Spanish, German, French, Italian, and Portuguese. It adds native punctuation and capitalization, and is more robust in reverberant environments. By compiling a single Conformer block and reusing its graph, IBM increased the model's throughput to 19,870 RTFx on the Open ASR Leaderboard, roughly 1.6x faster than its initial release of about 12,600 RTFx.

granite-speech-4.1-2b-plus is tuned for long recordings with multiple speakers, such as meetings and support calls. The updated version improves speaker attribution, timestamps, long-speech performance, and keyword biasing. Word diarization error rate drops 7%, average timestamp shift is reduced by more than half, long-form word error rate drops 24%, and biased-word errors fall 45% with 20-term keyword lists. The architecture, prompt format, and API are unchanged, making it a drop-in replacement.