Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
Source
huggingface.co
Date
Why it matters
New open speech models add keyword biasing for 2,000 terms, language-ID prompting and speaker attribution. This is useful if you build voice or transcription features into agents.
Key takeaways · AI-distilled
IBM's granite-speech-5.0-2b replaces 4.1-2b at the same size, adding language-ID prompting and keyword lists of up to 2,000 terms, and IBM ranks it #2 among open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → models on average English WER (5.06%, unofficial, Open ASR Leaderboard short-form sets).
The 470M Turbo CTC model went from English-only to seven languages with native punctuation, and compiling one Conformer block and reusing its graph raised it from about 12,600 to 19,870 RTFx, about 1.6x faster at the same accuracy.
granite-speech-4.1-2b-plus keeps the same architecture and API as a drop-in update for multi-speaker audio: average timestamp shift falls from 210.86 to 98.54, long-form WER drops 24%, and biased-keyword errors with 20-term lists fall 45%.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
