IBM published three updates to its Granite Speech collection on Hugging Face on October 21, 2026, according to a company blog post. The release includes a new model, granite-speech-5.0-2b, and revised versions of two existing models, granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. All three are released under an Apache 2.0 license with support in Transformers. According to IBM, granite-speech-5.0-2b transcribes English, French, German, Spanish, Portuguese, and Japanese, and translates to and from English in those languages, plus English to Italian and Mandarin. It was trained on 390,000 hours of public audio plus synthetic data. IBM states the model ranks #2 among top open-weight ASR models on average English word error rate (5.06%) on the Open ASR Leaderboard's public English short-form test sets as of October 21, 2026. The Turbo CTC model, previously English-only, is now multilingual and reported at 19,870 RTFx, which IBM describes as about 1.6x faster than its launch figure of roughly 12,600 RTFx at the same accuracy. The 4.1 2b-plus revision is tuned for meetings, support calls, and interviews with multiple speakers, and keeps an unchanged architecture and API as a drop-in replacement. IBM recommends pairing the models with Granite Guardian for production deployments.
- Three models released under Apache 2.0 on October 21, 2026, with native Transformers support
- granite-speech-5.0-2b transcribes six languages and translates across several, trained on 390,000 hours of audio
- IBM reports granite-speech-5.0-2b at 5.06% average English WER, ranked #2 on the Open ASR Leaderboard's open-weight list
- Turbo CTC reported at 19,870 RTFx, about 1.6x faster than its August launch, per IBM
What it means for you
These are speech-to-text models for turning recorded audio — meetings, support calls, interviews — into text, with translation and 'who said what' speaker labelling. They are free to use and run on your own hardware rather than a paid service, which matters if you handle recordings you can't send to an outside vendor. For most organizations this is plumbing: useful if transcription is a real bottleneck for you, irrelevant otherwise.
Try this
If you already transcribe calls or meetings, spend an hour mapping what actually happens to those transcripts afterward — who reads them, what decisions they feed. Fix that handoff before you swap any tool; a faster transcriber doesn't help if the output goes nowhere.
Who should care
Teams with a technical person who can self-host models, and operations in regulated or privacy-sensitive settings (healthcare, legal, support) that can't send call recordings to a cloud service.
Skip this if
You have no in-house technical capacity to deploy a model, your current transcription (Teams, Zoom, Otter, etc.) already works, or transcription simply isn't part of your work.
Sources: Hugging Face — read the original