Short version. IBM released granite-speech-5.0-2b on 20 October under Apache 2.0, along with refreshed Turbo CTC and 4.1 2B Plus weights. IBM's post says the 2B ranks second among open-weight speech models at 5.06% average English WER. When we pulled the live Open ASR Leaderboard data on 23 October, the same model sat at rank 10 overall with 5.68%, and it was no faster than its predecessor. The gains are real but smaller than the headline, and the punctuation numbers are weaker than the post implies.
You've got a meeting-transcription box running granite-speech-4.1-2b, and someone forwards the 5.0 announcement with "faster and more accurate" in the title. Do you swap it in? Before you do, it's worth reading the model card and the leaderboard next to the blog post, because the three don't say quite the same thing.
Second place, or tenth?
IBM's post has a chart saying the 2B ranks #2 among top open-weight ASR models on average English WER, 5.06%, averaged over the leaderboard's public English short-form sets. We can't reproduce that figure, and we don't think it's wrong. It's a different mix of test sets. The live main table on the Open ASR Leaderboard averages eleven columns by default, including URGENT24, Voice Arena Monsoon and two private sets, and there the model is listed at rank 10 of 57 with 5.68% and an RTFx of 544.5.
Among entries that link to a Hugging Face or GitHub page, only two score better on that table: Qwen3-ASR-1.7B at 5.40% (and 819 RTFx), and a preview called sophea/asr-k1 at 5.23%, whose link goes to a docs Space, so we can't say it's open. Cohere's transcribe model sits just behind at 5.77%. So "#2 open" holds on IBM's chosen sets and looks more like "#3 or so" on the default view. The leaderboard's private-data tab is less kind: rank 22, 8.47% average, 13.79% on conversational audio.
What "faster" and "7% better" cover
The title says faster. That's Turbo CTC, the 470M encoder, which IBM puts at 19,870 RTFx, about 1.6x its launch figure. The 2B itself isn't faster. On the live table the 5.0 2B shows 544.52 RTFx and 4.1 2B shows 544.94, and the average WER moves from 5.96% to 5.68%, a drop of about 4.7%. Same size, same speed, a bit better.
The 4.1 2B Plus table checks out against its own prose, with caveats. Speaker-attributed WER goes from 30.50 to 28.48 (6.6%), word diarization error from 4.40 to 4.10 (6.8%). Timestamp shift falls from 210.86 to 98.54 on all speech, which is the "more than half" claim, but only from 29.27 to 26.20 on short speech, about 10%. Keyword WER (11.79 to 6.48) and long-form WER (13.53 to 10.22) match the 45% and 24% in the text. Note that cpWER of 28.5 still means roughly one word in four wrong once you ask who said it.
Punctuation is where we'd slow down. The card reports a punctuation error rate of 23.20 on LibriSpeech clean and 21.22 on the other set, improved from 25.70 and 22.27, so about one mark in five is still wrong. It's worse than 4.1 on CV-EN (9.18 against 9.13) and CV-FR (11.43 against 11.00). The post says capitalization F1 ranges from 89.7 to 99.5, but 89.71 is the 4.1 figure. The 5.0 range on the card is 90.16 to 99.51. A tiny slip, though it's the kind that suggests nobody re-read the table.
What we'd do with it
Test it on your own audio. The licence is Apache 2.0, the base is granite-4.0-1b-base, and transformers 4.52.1 or later loads it. The card asks for mono 16 kHz audio and segments of about two minutes, so long recordings need voice activity detection in front. If you run the 4.1 2B Plus, pin the revision before pulling: IBM kept the name and swapped the weights underneath, which is convenient and also how a regression slips into a pipeline unnoticed.
I'd also keep the model at arm's length for English-only work where Qwen3-ASR-1.7B is available, since it scored better and ran faster on the same table. The 2B's case is languages: French, German, Spanish, Portuguese and Japanese in one open model, plus keyword lists up to 2,000 terms. We track its licence and size in the open-weight models tracker.
FAQ
Is Granite Speech 5.0 2B open source?
The weights are on Hugging Face under Apache 2.0, and the card lists granite-4.0-1b-base as the starting point and 390,000 hours of public audio plus synthetic data for training.
Which languages does it transcribe?
English, French, German, Spanish, Portuguese and Japanese. It translates to and from English in those languages, and English to Italian and Mandarin.
Is the 2B model faster than 4.1?
Not on the live Open ASR table, where both sit near 544 RTFx. The speed gain IBM reports is for the 470M Turbo CTC model.
Can it handle long recordings?
The card recommends segments of up to about two minutes and suggests splitting longer audio with a voice activity detector before decoding.
Sources. IBM Granite, Granite Speech 5.0 post, 21 October 2026; granite-speech-5.0-2b model card; Open ASR Leaderboard, read 23 October 2026.
