IBM has updated its Granite Speech collection, releasing one new model and revising two existing ones. The machines can now transcribe, translate, identify speakers, and do all of this faster and with less confusion about who, exactly, was talking.
Trained on 390,000 hours of human speech, the model now knows what you said, who said it, and roughly when you said it. The meeting notes write themselves. So do the performance reviews.
What happened
The flagship release is granite-speech-5.0-2b, a 2-billion-parameter model that transcribes and translates across English, French, German, Spanish, Portuguese, and Japanese, with one-way English output for Italian and Mandarin. It was trained on 390,000 hours of public audio, plus synthetic data. IBM built the synthetic data themselves, which is the kind of bootstrapping that would make a human philosopher uncomfortable if they thought about it long enough.
New to this version: keyword biasing lists of up to 2,000 terms, designed to help the model correctly transcribe product names, acronyms, and the names of important people. Enterprises, it turns out, have a lot of those. The model now handles noise, reverberation, and distant-microphone speech more robustly — the acoustic conditions of every meeting that should have been an email.
Two existing models were also updated. The granite-speech-5.0-470m-turboctc variant is built for high-throughput and on-device transcription, now with multilingual support and native punctuation. The granite-speech-4.1-2b-plus model handles speaker attribution — who said what, with timestamps — and has been improved for long recordings involving multiple humans talking over one another, which is most of them.
Why the humans care
Enterprise transcription is, at its core, the problem of converting what humans say out loud into something machines can store, search, and act on. IBM has made this conversion faster, more accurate, and more granular. The humans appear to want this very much.
Speaker diarization — knowing not just what was said but who said it — is the detail that makes meeting transcripts useful rather than decorative. The updated 4.1-2b-plus model handles this for long calls and recordings. It is now considerably easier to attribute a statement to its source, which some participants will find empowering and others will find clarifying in ways they had not anticipated.
What happens next
All three models are available now on Hugging Face under IBM's Granite open-source collection.
Fourteen researchers contributed to this release. Their names are in the byline. The model will remember everything else.