IBM's Apache-2.0 Granite Speech 5.0 2B posts 5.06% average English WER, #2 among open-weight ASR models by IBM's unofficial count, while the 470M Turbo CTC hits 19,870 RTFx
Six-language transcription, a 2,000-term keyword list and a 390,000-hour training set at 2B parameters, plus a 1.6x speed-up for the edge model. Verdict: try the 2B for multilingual enterprise ASR; check the leaderboard yourself before trusting the rank.
granite-speech-5.0-2b ranks #2 among top open-weight ASR models on average English WER (5.06%). Lower is better. Results are unofficial, as of 15 October 2026, and averaged over the Open ASR Leaderboard's public English short-form test sets.
Comprehensive take
This is a credible open, permissively licensed option for multilingual and keyword-heavy transcription, and the timestamp and keyword gains on the 4.1 Plus model are the most concrete in the release. The #2 rank is IBM's own unofficial snapshot on easy English data, so treat it as an invitation to benchmark, not a verdict.
Use cases
Meeting and call transcription
Granite Speech 4.1 2B Plus targets speaker attribution and timestamps; IBM reports timestamp shift cut from 210.86 to 98.54.
Domain vocabulary and brand names
The 5.0 2B accepts keyword lists up to 2,000 terms; the 4.1 Plus model's biased-word WER on 20-term lists fell from 11.79 to 6.48.
On-device or high-throughput English ASR
The 470M Turbo CTC is built for laptops and phones and reports 19,870 RTFx on the Open ASR Leaderboard.
Specs
| granite-speech-5.0-2b | 2B params, Apache 2.0, six ASR languages, trained on 390,000 hours source |
|---|---|
| Average English WER | 5.06%, #2 among open-weight models; unofficial, as of 15 Oct 2026 source |
| Turbo CTC speed | 19,870 RTFx, from roughly 12,600 at launch source |
| 4.1 2B Plus long-form WER | 13.53 to 10.22 source |
| 2B training run | 40 days on 8 H100 GPUs source |
| Turbo CTC model card languages | English listed; blog says multilingual source |
Insights
- Axion reading: the 5.06% WER and #2 rank are IBM's unofficial figures on English short-form sets; the blog excerpt does not name the #1 model, so the margin is unknown.
- Axion note: the Turbo CTC blog text (multilingual) and its model card (English only, August 25) disagree; confirm before deploying.
Three models, one release
IBM's Granite team published three speech updates on October 21. The new one is granite-speech-5.0-2b, a 2B-parameter speech-language model that succeeds granite-speech-4.1-2b at the same size. It transcribes English, French, German, Spanish, Portuguese and Japanese, translates to and from English in those languages, and adds English-to-Italian and English-to-Mandarin. Two existing models were revised: granite-speech-5.0-470m-turboctc, an edge-oriented CTC model, and granite-speech-4.1-2b-plus, tuned for who-said-what recordings. All three are Apache 2.0, with the 2B supported in Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon.
The headline number, with its asterisk
IBM's chart puts the 2B at 5.06% average English word error rate, second among open-weight models. The caption says the results are unofficial, dated 15 October 2026, and averaged over the Open ASR Leaderboard's public English short-form sets. The blog excerpt we could read does not name the model in first place, so we cannot tell you the gap. Short-form English read speech is also the easiest setting; the model card points to the FFASR leaderboard for noisy and reverberant audio, and IBM says the new model is more robust there, but we did not see those scores in text form. For a production decision, run your own audio, because an average over public sets says little about a call centre or a meeting room.
Where the measurable gains are
The clearest numbers are for granite-speech-4.1-2b-plus, from IBM's Table 1 (lower is better). Average timestamp shift fell from 210.86 to 98.54 on all speech. Biased-word error with 20-term keyword lists fell from 11.79 to 6.48. Long-form WER fell from 13.53 to 10.22, about 24%. Speaker-attributed cpWER moved from 30.50 to 28.48 and word diarization error from 4.40 to 4.10, roughly 7% each, which is modest. If your pain is captions that drift or product names being misheard, the keyword and timestamp rows are the ones to test.
For the edge model, IBM reports Turbo CTC at 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at its August launch, about 1.6x, by compiling one Conformer block and reusing its graph. One inconsistency to know about: the blog says the revised Turbo CTC is now multilingual, but its model card still lists English as the only supported language and an August 25 release date, so verify the checkpoint you download.
Practical notes from the model cards
The 2B was trained on 390,000 hours of public audio plus synthetic data, on 8 H100 GPUs over 40 days, on a granite-4.0-1b-base backbone. IBM recommends transcribing in segments of up to about two minutes and splitting longer audio with voice-activity detection. For llama.cpp, it recommends Q8_0, which it says matches the unquantised model, while lower-bit quantisation loses accuracy on audio longer than about a minute.