IBM's Granite team released three Granite Speech models on Hugging Face on Oct. 21, 2026, and each one is for a different speech job. Use granite-speech-5.0-2b for accurate transcription, speech translation and keyword biasing. Use granite-speech-5.0-470m-turboctc for fast or on-device transcription. Use granite-speech-4.1-2b-plus for meetings and multi-speaker calls that need speaker labels and word timestamps. IBM's announcement gives this same split. All three are Apache 2.0. Every performance figure below comes from IBM's own blog post and model cards, and no independent test of these releases was found.
Which model fits which job
| Job | Model | Why, per IBM |
|---|---|---|
| Accurate transcription, translation, domain terms | granite-speech-5.0-2b | Six ASR languages, translation to and from English, keyword lists up to 2,000 terms |
| High-throughput or on-device transcription | granite-speech-5.0-470m-turboctc | 470M-parameter CTC model, seven languages, aimed at laptops, smartphones and edge devices |
| Meetings, calls, captions with timing | granite-speech-4.1-2b-plus | Speaker labels, word-level timestamps, long-form evaluation up to 9 minutes |
granite-speech-5.0-2b: accuracy and translation
The model card gives Oct. 20, 2026 as the release date. It replaces granite-speech-4.1-2b and has the same parameter count.
- Transcription: English, French, German, Spanish, Portuguese and Japanese.
- Translation: Speech-to-text translation to and from English in those languages, plus English to Italian and English to Mandarin.
- Keyword biasing: Lists of up to 2,000 terms, for both transcription and translation. This helps with names, product terms and acronyms.
- Training: 390,000 hours of public audio, plus synthetic data for Japanese ASR, keyword biasing and translation (blog post).
- Runtimes: Transformers (version 4.52.1 or later), vLLM (offline mode and an OpenAI-compatible server), llama.cpp (GGUF builds, with Q8_0 recommended) and mlx-audio 0.5.9 or later on Apple Silicon.
Reported accuracy
The blog post says the model ranks #2 among top open-weight ASR models on average English word error rate (WER), at 5.06%. The chart caption calls the results "official, as of 21 October 2026," averaged over the Open ASR Leaderboard's public English short-form test sets. The model card also has a punctuation table that compares the model with 4.1-2b. Punctuation error rate falls on most test sets, for example LibriSpeech clean from 25.70 to 23.20. Capitalization F1 is mixed, though: it rises on LibriSpeech clean (89.71 to 90.16) and dips slightly on VoxPopuli (95.35 to 95.07).
Practical notes from the model card
- Transcribe audio in segments of up to about 2 minutes. Split longer recordings with voice activity detection (VAD) first.
- In llama.cpp, lower-bit quantizations lose accuracy on audio longer than about one minute.
- The Transformers example expects 16 kHz mono audio.
Reported changes in granite-speech-4.1-2b-plus
Which reported change matters most for your recordings? Sort by the size of the change.
| Measure | Reported change (%) | Applies to | Source |
|---|---|---|---|
| Speaker-attributed WER (cpWER), overall | 7 | Speaker-attributed transcripts | Source: Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and… (huggingface.co), Oct 21, 2026 |
| Word diarization error rate (WDER) | 7 | Speaker attribution | Source: Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and… (huggingface.co), Oct 21, 2026 |
| WER on long-form test sets | 24 | Long recordings | Source: Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and… (huggingface.co), Oct 21, 2026 |
| Errors on biased keywords | 45 | 20-term keyword lists | Source: Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and… (huggingface.co), Oct 21, 2026 |
All figures are IBM's self-reported results, and IBM's blog post does not name the test sets.
granite-speech-5.0-470m-turboctc: speed and edge devices
IBM first released TurboCTC as an English-only model on Aug. 25, 2026. The multilingual revision came out on Oct. 21, 2026. The English-only version is still available at revision 947f59a.
- Languages: English, French, German, Spanish, Portuguese, Italian and Japanese.
- Output: Transcripts with native punctuation and capitalization.
- Design: A non-autoregressive CTC model with greedy decoding, built on a 16-layer conformer encoder. The card describes its intended use as low-latency, high-throughput speech-to-text on laptops, smartphones and edge devices.
- Not covered: The card does not mention translation or keyword biasing. If you need either, use the 2B model.
Reported results
- Speed: The blog post says throughput rose to 19,870 RTFx on the Open ASR Leaderboard, up from about 12,600 at launch. IBM describes this as about 1.6x faster at the same accuracy level. The blog does not say whether these speed results are official.
- English: Average WER on Open ASR short-form tests is 5.51%, down from 5.78% for the English-only release (model card).
- Japanese: The card says the 470M model has a lower character error rate than granite-speech-5.0-2b on all three Japanese test sets: CSJ eval1, CommonVoice-17 and FLEURS. The numbers appear only in charts, so they are not repeated here.
granite-speech-4.1-2b-plus: meetings and multi-speaker audio
The model card gives April 28, 2026 as the release date, with an update on Oct. 19, 2026. IBM's blog calls the update "a drop-in replacement for the current model with improved weights." It keeps the same architecture, prompt format and API.
- Speaker attribution: Adds
[Speaker N]:labels before each speaker turn, numbered in order of appearance. - Word timestamps: Adds end-of-word tags in centiseconds (
[T:N]). The tags roll over every 10 seconds, so your code has to unwrap them. - Languages: English, French, German, Spanish and Portuguese. Japanese is not supported.
- Long-form: Evaluated on segments up to 9 minutes for ASR and speaker attribution, and up to 3.5 minutes for timestamps.
What IBM reports changed
From the blog post (previous release to update, lower is better):
| Measure | Before | After | IBM's description |
|---|---|---|---|
| Word diarization error rate (WDER) | 4.40 | 4.10 | 7% drop |
| Speaker-attributed WER (cpWER) | 30.50 | 28.48 | 7% drop overall |
| Long-form WER | 13.53 | 10.22 | 24% drop |
| Biased-word WER, 20-term lists | 11.79 | 6.48 | 45% fewer keyword errors |
| Average timestamp shift (AAS), all speech | 210.86 | 98.54 | More than halved |
The blog does not name the test sets behind these numbers. The model card's per-set tables show a more uneven picture:
- Average WER on the new Open ASR test sets went slightly up, from 6.01 to 6.10. Gigaspeech improved (9.02 to 8.70) and AMI got worse (7.88 to 9.04).
- WDER is mixed by dataset. It fell on AMI-SDM (14.6 to 13.9) and GALE (30.2 to 28.3), and rose on CALLHOME (2.2 to 3.0) and FISHER (0.9 to 1.0).
- Timestamp shift on long segments fell from 814.3 ms to 328.3 ms. The English average fell from 38.8 ms to 34.6 ms.
The gains are clearest for keyword biasing and timestamps. Plain transcription accuracy did not improve across the board. If your audio looks like AMI or CALLHOME, test the update on your own recordings before you switch.
Practical notes from the model card
- The model does not output punctuation or capitalization.
- Set
repetition_penalty=2.0. Without it, the model can loop, mostly on long audio and in speaker-attribution and timestamp modes. IBM's new leaderboard results usednum_beams=4with that penalty.
How far to trust the numbers
All figures come from IBM, the developer of these models. The 470M speed figures and the 4.1-2b-plus percentages do not name an official leaderboard result or any test sets. Both model cards warn that outputs can be inaccurate, biased or offensive, and IBM recommends pairing the models with Granite Guardian.
Limitations: this article relies only on IBM's own blog post and model cards, as published on Oct. 21, 2026, and no figure has been checked independently.
