The details
The new 2B model replaces Granite Speech 4.1 at the same size and copes better with background noise, echo and distant microphones. Users can name the input language in the prompt and supply lists of up to 2,000 terms, such as product names or acronyms, so they come out right. A third model, granite-speech-4.1-2b-plus, is a drop-in update for meetings and calls: IBM says errors in labelling who is speaking fell 7% and errors on long recordings fell 24%. All three use the Apache 2.0 licence and run in Transformers, and IBM suggests pairing them with its Granite Guardian models in production.
Why it matters
Speech-to-text is the first step in many AI products, from meeting notes to voice assistants. An openly licensed model that ranks second among open models on a public leaderboard lets teams run transcription on their own servers or devices and keep recordings private. The accuracy and speed figures come from IBM, so independent tests on messy, real-world audio will show how well it holds up.