IBM shipped three speech updates in the Granite Speech collection on October 21: a new granite-speech-5.0-2b, a revised granite-speech-5.0-470m-turboctc, and refreshed weights for granite-speech-4.1-2b-plus. All three are Apache 2.0. If you run transcription in-house, the useful part is that each model maps to a distinct job, so you can stop guessing which one to deploy.
Which of the three models to pick
| Model | Status | Pick it for |
|---|---|---|
granite-speech-5.0-2b |
New | High-accuracy multilingual transcription and translation, keyword biasing |
granite-speech-5.0-470m-turboctc |
Updated | High-throughput and on-device transcription |
granite-speech-4.1-2b-plus |
Updated | Meetings, calls, multi-speaker recordings |
IBM's own guidance is the same: accuracy, translation or keyword biasing goes to the 5.0 2B; raw speed or on-device goes to TurboCTC; multi-speaker audio goes to 4.1 2B Plus.
Granite Speech 5.0 2B: language prompts and 2,000-term keyword lists
The 5.0 2B is the successor to 4.1 2B at the same parameter count. It transcribes English, French, German, Spanish, Portuguese and Japanese. It translates to and from English in those languages, plus English to Italian and Mandarin. Training used 390,000 hours of public audio plus synthetic data for Japanese ASR, keyword biasing and translation.
What changed against 4.1:
- Better English short-form accuracy, plus more robustness to noise, reverberation and distant microphones.
- Language ID prompting: name the input language in the prompt to steer multilingual transcription.
- Keyword lists up to 2,000 terms, to bias toward product names, people and acronyms in every supported language.
- Punctuation and capitalization on request via the prompt. Capitalization F1 ranges from 89.7 to 99.5 across IBM's test sets.
- Better long-form and conversational translation.
IBM reports an average English WER of 5.06% on the Open ASR Leaderboard's public English short-form sets, ranking #2 among top open-weight ASR models as of 21 October 2026. That is IBM's chart, so check the leaderboard yourself before you commit.
Runtime support is broad: Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
TurboCTC: multilingual and about 1.6x faster
The August release was English-only. This revision keeps the 470M footprint and adds English, Japanese, Spanish, German, French, Italian and Portuguese. Transcripts now come with native punctuation and capitalization, so there is no separate punctuation or truecasing step. Training also targets reverberant audio, which should help in echoey rooms and far-field recordings.
Speed: compiling a single Conformer block and reusing its graph across the stack cut warmup time. TurboCTC now reaches 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at launch. That is about 1.6x faster at the same accuracy level, per IBM. There is also a WebGPU streaming demo for trying it in the browser.
4.1 2B Plus: a drop-in swap with measurably better diarization
This is the one I would update first if you process meetings or support calls. The name stays 4.1 because architecture, prompt format and API are unchanged. You get improved weights as a drop-in replacement.
| Metric (lower is better) | Previous | Updated |
|---|---|---|
| Timestamps, all speech (AAS) | 210.86 | 98.54 |
| Timestamps, short speech (AAS) | 29.27 | 26.20 |
| Biased-word WER, 20-term lists | 11.79 | 6.48 |
| Long-form WER | 13.53 | 10.22 |
| Speaker attribution (WDER) | 4.40 | 4.10 |
| Speaker-attributed cpWER | 30.50 | 28.48 |
The big movers are timestamps (average shift cut by more than half on all speech) and long-form WER (down 24%). Speaker attribution gains are smaller, about 7% on both WDER and cpWER. If your pain point is who said what, expect a modest improvement, not a transformation.
What I would do with it
The pitch is one model handling transcription, translation and speaker attribution instead of a chain of separate stages. Be careful with that framing. The source does not claim the 5.0 2B does diarization; speaker attribution lives in the 4.1 2B Plus. So you still pick a model per job, you just pick among fewer, Apache-licensed ones.
My plan: swap 4.1 2B Plus into an existing meeting pipeline, since the API is unchanged and the cost is a weights download. Then A/B the 5.0 2B with a domain keyword list against your current setup. Test the 2,000-term limit with your real vocabulary, not demo terms.
The catch: all numbers here are IBM's. The article points to model cards for full results, and says to pair the models with Granite Guardian for production deployments that need additional risk detection. Language coverage is also narrower than general-purpose ASR, with six languages for the 2B and seven for TurboCTC.
This follows the Granite Switch 4.1 post, which covered speech behind a single endpoint.
Sources: