IBM has released Granite Speech 5.0, a refresh of its open speech AI family that covers transcription, translation and multi-speaker recordings. The update adds one new model and improves two existing ones, all free to use under the Apache 2.0 license. If you build apps that turn audio into text, IBM Granite Speech now gives you three clear options depending on whether you care most about accuracy, speed or knowing who said what.
What IBM Granite Speech 5.0 Includes
The release is really three updates bundled together. Each model targets a different kind of audio work, so IBM is pitching them as a set rather than one model that does everything.
- granite-speech-5.0-2b: a new 2-billion-parameter model focused on accurate transcription and translation.
- granite-speech-5.0-470m-turboctc: a revised small model built for raw speed and on-device use.
- granite-speech-4.1-2b-plus: updated weights for meetings, support calls and interviews with several speakers.
All three ship with native support in the popular Transformers library. IBM says full test results, training details and usage examples are on each model card.
The New 2B Model Aims for Accuracy
The headline model, granite-speech-5.0-2b, replaces the earlier 4.1 version at the same size. It transcribes English, French, German, Spanish, Portuguese and Japanese, and it translates to and from English in those languages. It can also translate English into Italian and Mandarin.
IBM says it trained the model on 390,000 hours of public audio, plus synthetic data made for Japanese speech recognition, keyword biasing and translation. The company reports better handling of noisy rooms, echoes and speakers who sit far from the microphone. By IBM’s own unofficial count, the model ranks second among top open-weight speech recognition models on average English word error rate, at 5.06%.
Handy Controls for Real Apps
A few new features make the model easier to steer. You can name the input language in the prompt, feed it a keyword list of up to 2,000 terms so product names and acronyms come out right, and ask for punctuation and capitalization when you want ready-to-read text. It also runs in vLLM, llama.cpp and mlx-audio for Apple Silicon Macs.
Turbo CTC Gets Multilingual and Faster
When IBM introduced Turbo CTC in August, it only handled English. The revised 470-million-parameter version now transcribes seven languages: English, Japanese, Spanish, German, French, Italian and Portuguese. It also adds built-in punctuation and capitalization, so you skip a separate cleanup step.
Speed is the big story here. IBM says engineering changes pushed the model to a real-time factor of 19,870 on a public speech leaderboard, up from about 12,600 at launch, or roughly 1.6 times faster at the same accuracy. You can even try it in a web browser through a streaming demo IBM built on WebGPU.
Better Speaker Tracking in Granite Speech 4.1 Plus
The third model keeps its 4.1 name because its design, prompt format and API did not change. That makes it a drop-in swap for teams already using it, just with better weights. IBM tuned it for long recordings where several people talk, which is common in business settings.
| Area | Improvement IBM Reports |
|---|---|
| Speaker labeling errors | Down 7% |
| Speaker-attributed word errors | Down 7% overall |
| Timestamp drift | Cut by more than half |
| Errors on long recordings | Down 24% |
| Errors on biased keywords (20-term lists) | Down 45% |
Tighter timestamps help with captions and searching through recordings. Better speaker labels matter if you want an accurate record of a meeting or customer call.
Open speech models are getting good enough that many teams may no longer need a paid transcription API for everyday jobs.
What This Means for You
For most people, the impact will show up inside apps that use these models behind the scenes. Developers and IT teams get the more direct benefit, and the choice is fairly simple.
- Pick the 5.0 2B model if you need high accuracy, translation or custom vocabulary.
- Pick Turbo CTC if you need fast or on-device transcription, such as live captions.
- Pick 4.1 2B Plus if you process meetings or calls with multiple speakers.
- Test on your own audio first, since IBM’s benchmark numbers are its own and marked unofficial.
If you already rely on tools for automatic meeting notes, models like these are part of what keeps such features cheaper and more private. You can compare options in our AI meeting tools list. The release also joins a busy month for open models, following launches like Nvidia’s Nemotron 3.5 and new real-time voice models from ElevenLabs.
How It Fits the Open Model Race
IBM keeps positioning Granite as a business-friendly open family, and the permissive license is a big part of that. Companies can run the models on their own hardware, which keeps sensitive audio in-house. IBM also suggests pairing them with its Granite Guardian safety models for production use.
For more updates like this, browse our latest AI news page.
Sources
Primary source: IBM Granite’s Granite Speech 5.0 announcement.
Conclusion
Granite Speech 5.0 is a practical update rather than a flashy one, but it covers the three things teams ask for most from speech AI: accuracy, speed and clear speaker tracking. The new 2B model brings stronger transcription and translation, Turbo CTC gets much faster and learns six more languages, and the 4.1 Plus model gets better at multi-speaker recordings. Because everything is open under Apache 2.0, you can test all three without signing a contract. Try them on your own audio before switching, and you may find a capable free alternative to paid transcription services.
