IBM Releases Granite Speech 5.0 Under Apache 2.0

IBM’s newest speech model can be handed a list of up to 2,000 words to listen out for: product names, people and acronyms, in IBM’s examples.

Granite Speech 5.0 2B, released on Hugging Face on 21 October under the Apache 2.0 licence, is the headline of a three-model update from IBM’s Granite team. It transcribes English, French, German, Spanish, Portuguese and Japanese, and translates to and from English in those languages, plus from English into Italian and Mandarin. IBM also refreshed two models it released earlier: the fast Turbo CTC recogniser and the meeting-focused 4.1 2B Plus.

IBM puts the new model second on English accuracy

The 2B model replaces Granite Speech 4.1 2B at the same size. IBM says it was trained on 390,000 hours of public audio, plus synthetic data built for Japanese, keyword biasing and translation, and that it copes better with noise, echo and distant microphones. According to the model card, it was built by aligning an intermediate checkpoint of IBM’s Granite 4.0 1B language model to speech.

IBM’s chart puts it “#2 among top open-weight ASR models on average English WER (5.06%)”, meaning it got about five words in a hundred wrong on average across the Open ASR Leaderboard’s public English test sets. IBM labels those results “unofficial, as of 15 October 2026”. Users can also name the input language in the prompt, and ask for punctuation and capital letters. It runs in Transformers, vLLM, llama.cpp and Apple’s mlx-audio.

Turbo CTC learns six more languages and gets faster

Turbo CTC, a 470-million-parameter recogniser built for speed and on-device use, launched in August as an English-only model. IBM bundled it into Granite Switch 4.1 last week. The new version adds Japanese, Spanish, German, French, Italian and Portuguese, and produces punctuated text without a separate step.

IBM says a compiler change pushed the model to 19,870 RTFx on the Open ASR Leaderboard, against roughly 12,600 at launch, about 1.6 times faster. RTFx measures how many seconds of audio a model processes per second of compute. In August, IBM said 12,600 RTFx on an NVIDIA H200 meant transcribing more than 3.5 hours of speech in one second.

The third model, Granite Speech 4.1 2B Plus, keeps its name because IBM says it is a drop-in replacement with new weights. IBM’s figures show its word diarisation error rate, which counts words credited to the wrong speaker, falling from 4.40 to 4.10. Error on long recordings drops from 13.53 to 10.22, and errors on boosted keywords from 11.79 to 6.48. IBM says the average timestamp drift was more than halved, from 210.86 to 98.54.

All three models are on Hugging Face under Apache 2.0, which allows commercial use. Every figure so far comes from IBM. The model card already cites official Open ASR Leaderboard results from 20 October, and the public leaderboard is where to check whether the 2B model keeps its place among open models as new entries arrive.