IBM releases Granite Speech 5.0
IBM's Granite team published granite-speech-5.0-2b, a 2-billion-parameter speech-language model that transcribes six languages and translates to and from English, together with revised versions of two older models, the 470-million-parameter Turbo CTC and granite-speech-4.1-2b-plus. The model card gives 20 October 2026 as the release date, the announcement on Hugging Face is dated 21 October, and all three models are under Apache 2.0.
Why it matters
An enterprise-oriented speech model with an open licence and support in Transformers, vLLM, llama.cpp and mlx-audio joins the open models that can be run locally. The revision of Turbo CTC from English-only to seven languages shows the family growing in coverage. How good the new model is rests on IBM's own reading of the Open ASR Leaderboard, which moves.
What the pages say. Blog post of 21 October 2026 on Hugging Face (an enterprise article by George Saon and colleagues) and the model card of granite-speech-5.0-2b, which says "Release Date : October 20, 2026" and "License: Apache 2.0". The 2B model is the successor of granite-speech-4.1-2b at the same size; it transcribes English, French, German, Spanish, Portuguese and Japanese and translates to and from English in those languages, plus English to Italian and Mandarin, with language-ID prompting, 2,000-term keyword biasing and formatted output on request. Turbo CTC, introduced in August as an English-only encoder of 470 million parameters, is revised to transcribe English, Japanese, Spanish, German, French, Italian and Portuguese with punctuation and capitalisation. granite-speech-4.1-2b-plus keeps its name because architecture, prompt format and API are unchanged; it improves speaker attribution, timestamps, keyword biasing and long recordings. Dates. The 2B repository was created on 7 October 2026 and its checkpoint uploaded that day; the card's 20 October is the only statement of release, and the record keeps it as the day of the event and 21 October as the day of the announcement. The two revised models are older repositories (April and August 2026) updated on 21 October. What the record does not claim. The post says granite-speech-5.0-2b "ranks #2 among top open-weight ASR models on average English WER" on the Open ASR Leaderboard "as of 21 October 2026"; that is IBM's reading of a changing board and no figure from it is repeated here. No independent test was read.