IBM has released Granite Speech 5.0, a new open speech recognition model that transcribes six languages and translates between them and English, alongside updated versions of two older Granite speech models. All three were published on Hugging Face on 21 October 2026 under the Apache 2.0 licence, which means anyone can download, modify and use them commercially for free.
The headline model, granite-speech-5.0-2b, ranked second among open-weight models on the Open ASR Leaderboard for English short-form audio as of release day, according to IBM's Hugging Face post. Here is what each model does, who it suits, and where it falls short.
What Granite Speech 5.0 can do
The new 2-billion-parameter model replaces granite-speech-4.1-2b at the same size. IBM lists these capabilities:
- Transcription in English, French, German, Spanish, Portuguese and Japanese.
- Translation to and from English for those languages, plus English into Italian and Mandarin.
- Keyword lists of up to 2,000 terms, so names, product codes or medical words are spelled the way you want.
- Language identification by prompt, and punctuation and capital letters on request.
IBM says it trained the model on 390,000 hours of public audio, plus synthetic data for Japanese, keyword biasing and translation. It reports an average English word error rate of 5.06%, meaning roughly five words in every hundred come out wrong on the leaderboard's test sets. IBM also says the model copes better with background noise, echo and distant microphones than its predecessor.
The fast model: 470M TurboCTC
The second release is an update to granite-speech-5.0-470m-turboctc, a much smaller model built for speed and for running on devices. When it launched in August it handled English only. It now covers seven languages: English, Japanese, Spanish, German, French, Italian and Portuguese.
The speed figure is the striking one. IBM reports an RTFx of 19,870 on the Open ASR Leaderboard, up from about 12,600 at launch. RTFx measures how many seconds of audio a model processes per second of compute, so in plain terms it can work through hours of recordings in moments on suitable hardware. IBM says accuracy stayed at the same level while speed rose about 1.6 times.
It now adds punctuation and capitals itself, with no separate clean-up step, and IBM has put up a streaming demo that runs in a web browser using WebGPU.
The meeting model: 4.1-2b-plus
The third model, granite-speech-4.1-2b-plus, is aimed at calls, meetings and other recordings with several speakers. IBM kept the architecture, prompt format and API the same, so developers already using it can swap in the new weights without changing code. The improvements IBM reports:
- Long-form word error rate fell from 13.53% to 10.22%, about a quarter fewer mistakes.
- Errors on keywords from a 20-term list dropped from 11.79% to 6.48%, close to half.
- Speaker attribution error (WDER) improved from 4.40 to 4.10.
- Timestamp drift across all speech fell from 210.86 to 98.54 on IBM's measure.
How to use Granite Speech
All three models are on Hugging Face. IBM says they work natively with the Transformers library and are supported by vLLM for serving, llama.cpp for running on ordinary computers, and mlx-audio for Apple Silicon Macs. For production use, IBM suggests pairing them with Granite Guardian, its separate risk-detection model.
IBM's post does not list minimum hardware requirements. As a rough guide, the 470M model is the one to try first on a laptop or phone, while the 2B models will run more comfortably with a GPU.
What it means for users in Africa
Because the models are free and downloadable, there is no sign-up, regional block or subscription to worry about. A developer in Lagos or Nairobi can download them on the same terms as anyone else, and run them offline once downloaded, which helps where data is expensive.
The language list is the limitation. None of the three models supports an African language such as Swahili, Hausa, Yoruba, Amharic or Zulu. French and Portuguese coverage helps in francophone and lusophone countries, and English helps everywhere else, but accuracy on African-accented English is not reported separately in IBM's numbers. Anyone building a product for local users should test on their own recordings before relying on the leaderboard figures.
For the other side of this, see our explainer on Gradium's Phonon text-to-speech model, which turns text into speech on small devices.
Who should try it
- Journalists and researchers who transcribe interviews and want a free, private option.
- Teams building call-centre or meeting tools that need speaker labels.
- Developers who need fast batch transcription of large audio archives.
IBM ended its announcement with a simple request to users: "Give them a try and let us know what you think!"
