IBM Releases Granite Speech 5.0 2B Under Apache 2.0 and Updates Two Existing Granite Speech Models
IBM published granite-speech-5.0-2b and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus, all Apache 2.0, with IBM-reported leaderboard and error-rate figures.
Overview
IBM has published three updates to its Granite Speech collection on Hugging Face: a new speech-language model, granite-speech-5.0-2b, and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus, according to IBM’s Hugging Face blog post, dated October 21, 2026. The post says the models are “all released under an Apache 2.0 license with native support in Transformers.” All performance figures below are IBM’s own, as published in that post and the model card, and were read by the Herald on October 27, 2026. The sources reviewed contain no independent testing of them.
What IBM Published
Granite Speech 5.0 2B
The post describes granite-speech-5.0-2b as the successor to granite-speech-4.1-2b at the same parameter count. IBM says 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.
The listed changes compared with the 4.1 2B model are language-ID prompting, keyword lists of up to 2,000 terms, punctuation and capitalization on request, and better long-form and conversational translation. IBM reports that capitalization F1 “ranges from 89.7 to 99.5 across our test sets.” The post also lists support in vLLM, llama.cpp and mlx-audio for Apple Silicon. The model card lists a release date of October 20, 2026 and the Apache 2.0 license, and says: “For best results, transcribe audio in segments of up to ~2 minutes; split longer recordings (e.g. with VAD) before decoding.”
IBM’s chart caption states that the model “ranks #2 among top open-weight ASR models on average English WER (5.06%)”, with results “official, as of 21 October 2026” and averaged over the Open ASR Leaderboard’s public English short-form test sets. This is a ranking reported by IBM, not an independent evaluation.
Granite Speech 5.0 470M Turbo CTC
IBM writes that the Turbo CTC model, introduced in August as an English-only, 470M-parameter encoder, now transcribes English, Japanese, Spanish, German, French, Italian and Portuguese, produces native punctuation and capitalization, and is more robust to reverberant and far-field audio. The post says compiling a single Conformer block and reusing its graph pushed the model to 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch: “about 1.6x faster, at the same accuracy level.”
Granite Speech 4.1 2B Plus
The post says the revised granite-speech-4.1-2b-plus keeps its name because the “architecture, prompt format, and API are unchanged,” making it a drop-in replacement with improved weights. IBM’s table of previous-release versus updated-model error rates (lower is better) reports:
- Average timestamp shift (AAS), all speech: 210.86 to 98.54
- Biased-word WER with 20-term keyword lists: 11.79 to 6.48
- WER on long-form test sets: 13.53 to 10.22
- Word diarization error rate (WDER): 4.40 to 4.10
- Speaker-attributed WER (cpWER): 30.50 to 28.48
What We Don’t Know
The sources reviewed are IBM’s own blog post and model card. They do not include third-party replications of the leaderboard ranking, the speed figure or the error-rate changes, and the blog post does not describe how the leaderboard figures for the other two models were produced beyond naming the Open ASR Leaderboard. The post also recommends pairing the models with Granite Guardian “for production deployments that need additional risk detection.”