IBM Granite Speech 5.0: open ASR at 5.06% WER, Apache 2.0

IBM published three Granite Speech updates on 21 October 2026 on the Hugging Face blog: a new model, granite-speech-5.0-2b, and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. IBM says all three are released under the Apache 2.0 licence with support in Transformers, and that the 5.0 2B also runs in vLLM, llama.cpp and mlx-audio for Apple Silicon.

The new 2B model

granite-speech-5.0-2b succeeds the 4.1 2B at the same parameter count. IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese, translates to and from English in those languages, and also translates English to Italian and Mandarin. IBM says it was trained on 390,000 hours of public audio plus synthetic data for Japanese transcription, keyword biasing and translation.

• Accuracy: IBM reports an average English WER of 5.06% across the Open ASR Leaderboard's public English short-form test sets, which it says ranks #2 among open-weight models as of 21 October 2026. Lower is better.
• Custom vocabulary: keyword lists of up to 2,000 terms bias output toward product names, people and acronyms in every supported language.
• Formatting and language control: punctuation and capitalisation on request (IBM reports capitalisation F1 of 89.7 to 99.5 across its test sets), and a prompt that names the input language.
• Acoustics: IBM says it handles noise, reverberation and distant-microphone speech better than 4.1 2B.

The two updated models

The 470M Turbo CTC model, built for speed and on-device use, is now multilingual (English, Japanese, Spanish, German, French, Italian and Portuguese) and emits punctuation and capitalisation natively. IBM says compiling a single Conformer block and reusing its graph lifted it to 19,870 RTFx (real-time factor, audio seconds processed per second) from roughly 12,600 at launch, about 1.6 times faster at the same accuracy level.

The 4.1 2B Plus keeps its name because the architecture, prompt format and API are unchanged, so IBM calls it a drop-in replacement with better weights. It targets multi-speaker meetings and support calls. IBM's own comparison against the previous release:

• Long-form WER: 13.53 down to 10.22.
• Biased-keyword WER with 20-term lists: 11.79 down to 6.48.
• Timestamp shift (average, all speech): 210.86 down to 98.54.
• Speaker-attributed WER (cpWER): 30.50 down to 28.48; word diarization error rate: 4.40 down to 4.10.

What the post leaves out

Every figure comes from IBM. The leaderboard ranking uses public English short-form test sets, which are cleaner than most call-centre audio, and IBM gives no results for accented, telephone-band or heavily overlapping speech. The post gives no cost per audio hour, no hardware requirements for the 2B model and no word error rates for the non-English languages. IBM recommends pairing the models with its Granite Guardian safety models for production use; it does not say what that adds in latency or cost.

What to do

• If you transcribe calls or meetings today, pull a sample of 50 to 100 of your own recordings and compare WER on the 4.1 2B Plus and the 5.0 2B against your current service.
• Build a keyword list from your product names, customer names and acronyms and test it at 20 terms and at the full 2,000.
• Check the model cards for per-language results and the Apache 2.0 terms before any production use.