Glonce

IBM ships Granite Speech 5.0 2B, #2 open-weight model on English ASR

modelsThe Glonce editorsHugging Face🔥62Glonce rating62Impact55Novelty50Relevance100Surprise25Players55Credibility85
IBM ships Granite Speech 5.0 2B, #2 open-weight model on English ASR

IBM published three updates to its Granite Speech collection on the Hugging Face blog: a new speech-language model, granite-speech-5.0-2b, and revised versions of two models the community already uses, granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. IBM says together they cover enterprise speech work: accurate transcription and translation, very fast transcription at the edge, and recordings where who spoke and when matter as much as what was said. All three are released under Apache 2.0 with native support in Transformers.

Granite Speech 5.0 2B is the successor to granite-speech-4.1-2b at the same parameter count. It transcribes English, French, German, Spanish, Portuguese and Japanese, and translates to and from English in those languages, plus English to Italian and Mandarin. It was trained on 390,000 hours of public audio, plus synthetic data built for Japanese ASR, keyword biasing and translation. Compared with 4.1 2B, IBM lists five changes: higher accuracy on English short-form benchmarks with better handling of noise, reverberation and distant-microphone speech; language ID prompting, where the input language is named in the prompt to steer multilingual transcription; better long-form and conversational translation; keyword lists of up to 2,000 terms to bias transcription and translation toward product names, people and acronyms in every supported language; and punctuation and capitalization on request via the prompt, with capitalization F1 ranging from 89.7 to 99.5 across IBM's test sets. Support covers Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon. IBM's chart says the model ranks #2 among top open-weight ASR models on average English WER, at 5.06% (lower is better), official as of 21 October 2026 and averaged over the Open ASR Leaderboard's public English short-form test sets.

Granite Speech 5.0 Turbo CTC was introduced in August as an English-only, 470M-parameter encoder built for speed. The revision keeps that footprint and adds multilingual transcription (English, Japanese, Spanish, German, French, Italian and Portuguese), native punctuation and capitalization with no separate punctuation or truecasing step, and more robustness in echoey rooms and far-field recordings thanks to training that targets reverberant audio. It is also faster: by compiling a single Conformer block and reusing its graph across the stack, IBM cut warmup time and pushed Turbo CTC to 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch. That is about 1.6x faster at the same accuracy level. A second chart plots average WER against RTFx on nine public English short-form test sets, as of October 2026.

Granite Speech 4.1 2B Plus is tuned for meetings, support calls and long multi-speaker interviews. It keeps the 4.1 name because architecture, prompt format and API are unchanged, so it is a drop-in replacement with improved weights. IBM reports four gains: word diarization error rate (WDER) drops 7% and speaker-attributed WER (cpWER) drops 7% overall; average timestamp shift (AAS) is cut by more than half; WER on long-form test sets drops 24%; and errors on biased keywords fall 45% with 20-term keyword lists. All table metrics are error rates, so lower is better.

IBM's pick guide: use granite-speech-5.0-2b for high transcription accuracy, translation or keyword biasing; use granite-speech-5.0-470m-turboctc for raw speed or on-device transcription, which can be tried in the browser through a WebGPU streaming demo; use granite-speech-4.1-2b-plus for meetings and multi-speaker calls. IBM recommends pairing the models with Granite Guardian for production deployments that need additional risk detection.

Key facts

  • granite-speech-5.0-2b, successor to 4.1 2B at the same size, ranks #2 among top open-weight ASR models on average English WER (5.06%) on the Open ASR Leaderboard's public English short-form sets, as of 21 October 2026.
  • The 5.0 2B model was trained on 390,000 hours of public audio plus synthetic data, covers six transcription languages, and supports keyword lists of up to 2,000 terms.
  • Turbo CTC (470M parameters) is now multilingual across seven languages, outputs punctuation and capitalization natively, and reaches 19,870 RTFx, up from roughly 12,600 and about 1.6x faster at the same accuracy level.
  • granite-speech-4.1-2b-plus, a drop-in weight update, cuts WDER 7%, cpWER 7%, long-form WER 24% and keyword errors 45% (20-term lists), and more than halves timestamp shift.
  • All three models are Apache 2.0 with native Transformers support; the 5.0 2B model also works with vLLM, llama.cpp and mlx-audio.

Why it matters

Open-weight speech models with a permissive licence are a practical base for transcription products, and IBM puts the new 2B model second among top open-weight models on the Open ASR Leaderboard's average English WER (5.06%). The update also targets enterprise pain points: long recordings, multiple speakers, product-name vocabulary and speed on edge devices. Turbo CTC also goes from English-only to seven languages while gaining about 1.6x in speed (19,870 versus roughly 12,600 RTFx).

Who it affects

Developers building transcription, captioning, meeting-notes and call-analysis tools, especially teams that need open weights under Apache 2.0. Edge and on-device teams are the audience for Turbo CTC. Teams processing multi-speaker meetings, support calls and interviews are the audience for the 4.1 2B Plus revision. Apple Silicon users get mlx-audio support for the 5.0 2B model.

How to use it

IBM's guide: pick granite-speech-5.0-2b for accuracy, translation or keyword biasing; granite-speech-5.0-470m-turboctc for speed or on-device work, with a WebGPU streaming demo to try it in the browser; granite-speech-4.1-2b-plus for meetings and multi-speaker calls. In the 5.0 2B model, name the input language in the prompt to steer multilingual transcription, and ask for formatted output in the prompt if you want punctuation and capitalization. Keyword lists of up to 2,000 terms bias output toward names and acronyms. The 4.1 2B Plus revision keeps its architecture, prompt format and API, so it replaces the current model with no code changes. All three models are Apache 2.0 with native Transformers support; full evaluation results, architecture, training data and usage examples are on each model card.

How solid is it

All figures are IBM's own, from its blog post on Hugging Face. The #2 ranking and 5.06% WER are stated as official as of 21 October 2026 on the Open ASR Leaderboard's public English short-form test sets, so the standing can change. The post does not name which models rank #1 or ranks 3 and below. For 4.1 2B Plus, only relative reductions are given (7%, more than half, 24%, 45%), with no absolute error rates in the text. The Turbo CTC accuracy level is not quantified beyond "the same accuracy level", and no hardware details are given for the RTFx measurements.

Risks and caveats

The leaderboard ranking covers English short-form audio only and is an average over the leaderboard's public test sets; it says little about the other languages, long recordings or noisy conditions. Capitalization F1 spans 89.7 to 99.5 across IBM's test sets, so formatting quality varies by data. IBM itself recommends pairing the models with Granite Guardian for production deployments that need additional risk detection. The 4.1 2B Plus improvements are reported as percentage drops without absolute values in the post text, so the size of the real-world gain depends on your own audio.