IBM Upgrades Its Granite Speech AI Line With Faster Transcription, Better Translation, and Smarter Meeting Notes

Key points

• IBM's new granite-speech-5.0-2b scored an average 5.06% word error rate on English short-form benchmarks as of 15 October 2026, ranking second among open-weight speech AI models on the Open ASR Leaderboard.
• The speed-focused granite-speech-5.0-470m-turboctc now transcribes at 19,870 RTFx, roughly 1.6 times faster than when IBM launched it in August 2026.
• The meeting-focused granite-speech-4.1-2b-plus cuts average timestamp error by more than half and drops word error on long recordings by 24%.
• All three models are free to use under an Apache 2.0 open licence, meaning businesses can run them without royalty fees or vendor lock-in.

IBM shipped three updates to its Granite Speech family this week. Which version do you actually need?

The headline arrival is granite-speech-5.0-2b, a 2-billion-parameter model (parameter count is a rough measure of model size and capability) that transcribes six languages and translates between them and English. IBM trained it on 390,000 hours of public audio, and the results placed it second on the Open ASR Leaderboard, a public ranking of speech-recognition accuracy, with an average error rate of 5.06% on English tests. Every percentage point of error is a word your transcription software gets wrong.

One genuinely useful new trick: you can now tell the model which language you're speaking before it starts, rather than letting it guess. For anyone switching between English and Japanese mid-meeting, that's a real fix. Keyword lists of up to 2,000 terms let product names, staff names and acronyms get priority in the transcript rather than turning into garbled nonsense.

The speed champion is granite-speech-5.0-470m-turboctc. At 470 million parameters it's the smallest model in the group, built to run fast on a phone or a laptop without needing a server. IBM says it now processes audio at 19,870 RTFx, where RTFx measures how many times faster than real time the model runs. The August launch ran at roughly 12,600 RTFx; this version is about 1.6 times faster at the same accuracy level. Seven languages are now supported, and punctuation and capitalisation arrive automatically, so the raw transcript comes out ready to read.

The third update, granite-speech-4.1-2b-plus, targets the recordings enterprises actually deal with: long meetings and support calls where knowing who said what, and exactly when, matters as much as the words themselves. IBM kept the model architecture identical to the previous version so that companies already using it can swap in the new weights without changing any code. Average timestamp error dropped by more than half, from 210 milliseconds to 98 milliseconds on full-length recordings. Word error on long-form audio fell 24%. Errors on custom keyword lists fell 45%.

All three models ship under Apache 2.0 licence and run in standard developer frameworks. IBM published the numbers on Hugging Face, where the model cards carry full benchmark details.

We've been tracking this corner of the market closely: Apple's on-device compression work we reported in September showed how fast the efficiency race is moving, and IBM's numbers this week land in the same conversation. The real story isn't the benchmark. Open-licensed enterprise speech AI is now accurate enough that the argument for expensive proprietary systems gets harder to make every quarter.

Which model should a business actually use?

For most teams, the choice maps directly to the job. Need accurate transcripts or translations across languages? Use the 2B model. Need something that runs fast on a laptop or phone? Use Turbo CTC. Running meeting recordings where you need to know who said what? Use the 4.1 Plus update.

IBM recommends pairing any of them with its Granite Guardian safety layer for production deployments where you need automated risk checks on the output.

Common questions

Do I need expensive hardware to run these models?

Turbo CTC is specifically designed for on-device use, meaning an ordinary laptop or a modern phone can handle it. Both 2B models need more power but are well within reach of standard business servers.

Is my audio data kept private if I use these?

Because the models run locally rather than sending audio to a remote server, your recordings stay on your own hardware. That's one of the main reasons enterprises choose on-device AI for sensitive calls and meetings.