IBM voice model is number 2 on an open English test

IBM released three updates to its Granite Speech models: a new model, granite-speech-5.0-2b, and new versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. IBM says that the new model is number 2 of the open-weight voice-to-text models, with an average English word error rate of 5.06%. The result is unofficial. It is from 15 October 2026 and uses open test sets.

What is new

• The 2B model transcribes six languages and translates to and from English. It accepts keyword lists of 2,000 terms at the maximum.
• The 470M model is for fast work on a device. It has seven languages.
• The 4.1 plus model is for long meetings and calls with more than one speaker.

The numbers

• The speed of the 470M model is about 1.6 times as fast, 19,870 on the speed score, with the same level of precision.
• On long recordings, errors drop 24%.
• Errors on keywords drop 45%, and errors in speaker labels drop 7%.

What is not known

The results are from IBM and are unofficial. IBM gives no test by a different group. IBM tells users to add its Granite Guardian tool for risk detection in production.

Insight: IBM and Mistral compete on the same voice features

Mistral launched Voxtral 3 on 20 October 2026 with speaker names and term lists, but gave no scores. IBM adds the same two features, with scores and an Apache 2.0 license. This reads as a race to be the default voice layer, where open terms and open scores maybe get trust. Test the two on your own recordings.