IBM has refreshed its open speech stack. In a Hugging Face post dated October 21, 2026, the Granite team introduced a new speech-language model, granite-speech-5.0-2b, and revised versions of two existing models: granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. All three are released under Apache 2.0, with support for Transformers, and the 2B also runs in vLLM, llama.cpp and mlx-audio.
The update targets three jobs that enterprises actually have: accurate multilingual transcription and translation, fast on-device transcription, and meeting recordings where speaker and timing matter. We summarize each model, flag which numbers are IBM's own, and give a quick chooser. It follows IBM's Granite Switch 4.1 release, which put chat, decisions and speech behind one endpoint.
TL;DR: which Granite Speech model to use
| Model | Status | Best for | Headline claim |
|---|---|---|---|
| granite-speech-5.0-2b | New | High-accuracy multilingual transcription and translation | 5.06% average English WER, #2 among open-weight ASR models (IBM) |
| granite-speech-5.0-470m-turboctc | Updated | High-throughput and on-device transcription | 19,870 RTFx, about 1.6x faster than at launch |
| granite-speech-4.1-2b-plus | Updated | Meetings and multi-speaker calls | Long-form WER down 24%, timestamps shift cut by more than half |
Granite Speech 5.0 2B: the accuracy model
granite-speech-5.0-2b succeeds 4.1 2B at the same parameter count. IBM says it was trained on 390,000 hours of public audio plus synthetic data built for Japanese ASR, keyword biasing and translation. New capabilities compared with 4.1:
- Better English ASR and robustness. Higher accuracy on English short-form benchmarks and better handling of noise, reverberation and distant-microphone speech.
- Language ID prompting. Name the input language in the prompt to steer multilingual transcription.
- Better long-form and conversational translation. Translation holds up for longer, more natural speech instead of only short read sentences.
- Keyword lists up to 2,000 terms. Bias output toward product names, people and acronyms in every supported language.
- Formatting on request. Ask for punctuation and capitalization in the prompt; IBM reports capitalization F1 between 89.7 and 99.5 across its test sets.
Languages for transcription are English, French, German, Spanish, Portuguese and Japanese, with translation to and from English in those, plus English to Italian and Mandarin.
The leaderboard claim, and how to read it
IBM's chart says the model ranks second among top open-weight ASR models on average English word error rate, 5.06%, averaged over the Open ASR Leaderboard's public English short-form test sets, official as of October 21, 2026. Leaderboards move daily, English short-form sets are only one slice of real audio, and word error rate says nothing about latency or cost. Test it on your own recordings, especially if your audio is noisy, accented or domain-specific.
Turbo CTC: faster, now multilingual
When IBM introduced Turbo CTC in August it was an English-only 470M-parameter encoder built for speed. The revision keeps that size and adds multilingual transcription (English, Japanese, Spanish, German, French, Italian, Portuguese), native punctuation and capitalization with no separate truecasing step, and training that targets reverberant audio for echoey rooms and far-field recordings.
On speed, IBM compiled a single Conformer block and reused its graph across the stack, cutting warmup time and lifting throughput to 19,870 RTFx from roughly 12,600, about 1.6x, at the same accuracy level. An in-browser WebGPU streaming demo is available. If you care about local, private transcription, compare it with tools we have covered such as Tether's TranscribePsy and Cactus Whistle.
4.1 2B Plus: who said what, and when
The refreshed granite-speech-4.1-2b-plus keeps its name because the architecture, prompt format and API are unchanged, making it a drop-in swap with improved weights. IBM's table, where all metrics are error rates and lower is better:
| Capability | Metric | Previous | Updated |
|---|---|---|---|
| Timestamps, all speech | Average timestamp shift | 210.86 | 98.54 |
| Timestamps, short speech | Average timestamp shift | 29.27 | 26.20 |
| Keyword biasing | Biased-word WER, 20-term lists | 11.79 | 6.48 |
| Long-form speech | WER | 13.53 | 10.22 |
| Speaker attribution | WDER | 4.40 | 4.10 |
| Speaker-attributed transcripts | cpWER | 30.50 | 28.48 |
The percentage gains IBM cites: speaker-attribution error falls 7%, long-form WER falls 24%, and biased-keyword errors fall 45% with 20-term lists. Speaker attribution remains the hardest part of meeting transcription, and a 4.10 diarization error rate is a small step, so the practical gain here is mostly better timestamps and long-form accuracy.
How it compares with other speech options
Speech recognition is crowded. Closed APIs such as Gemini 3.5 Transcribe and Microsoft's MAI-Transcribe 2 compete on streaming and managed convenience; Granite competes on being open, permissively licensed and self-hostable. Our earlier Hugging Face speech-to-speech voice agent guide shows how an open ASR model plugs into an agent pipeline. IBM also recommends pairing the models with Granite Guardian for production deployments that need additional risk detection.
The metrics in IBM's tables, explained
Speech papers use abbreviations that make results hard to compare, so here is what each number in IBM's post measures.
- WER (word error rate) counts substitutions, deletions and insertions against a reference transcript, divided by the number of reference words. A 5.06% average means roughly five errors per hundred words on the leaderboard's English short-form sets.
- RTFx (inverse real-time factor) is how many seconds of audio the system processes per second of compute. At 19,870 RTFx, an hour of audio would take under a fifth of a second of compute on the leaderboard's setup, which is why the figure matters for batch jobs and call-center archives rather than for live captions.
- WDER (word diarization error rate) measures how often a correctly recognized word is assigned to the wrong speaker.
- cpWER (concatenated minimum-permutation WER) scores a speaker-attributed transcript, penalizing both word errors and wrong speaker labels.
- AAS (average timestamp shift) is the average gap between predicted and reference word timings; IBM's table reports it falling from 210.86 to 98.54 for all speech, which makes captions and search-to-audio jumps tighter.
- Biased-word WER is the error rate on the specific terms you supplied in a keyword list.
When comparing vendors, make sure the same metric and the same test sets are used. A model that looks excellent on clean read speech can lose badly on a call recorded on a speakerphone.
Practical playbooks
Customer-call analytics. Use granite-speech-4.1-2b-plus for speaker attribution and timestamps, add a keyword list of product names and plan names, then feed transcripts to your analytics or an LLM for summaries. Keep a human-reviewed sample to monitor drift, since speaker labels are the weakest part.
Multilingual support desk. Use granite-speech-5.0-2b with language ID prompting when you know the caller's language, and ask for formatted output so downstream search works on punctuated text. Remember that translation is listed only to and from English in the supported languages, plus English to Italian and Mandarin.
On-device dictation or edge capture. Use Turbo CTC for its footprint and speed. Native punctuation means you do not need a second model, and the WebGPU demo is a quick way to judge latency on user hardware before you commit.
Compliance and redaction pipelines. Transcribe first, then redact. Pair the transcript with a policy model; IBM suggests Granite Guardian for risk detection in production deployments.
Limits to keep in mind
IBM's own numbers are strong, but three caveats apply. First, all reported figures are vendor-run: they come from IBM's post and the public leaderboard, and independent replication on messy real-world audio is what ultimately matters. Second, the language lists are short compared with some commercial services, so teams that need Hindi, Arabic or other languages will not find them here. Third, the 4.1-2b-plus gains are modest in some rows, for example diarization error moving from 4.40 to 4.10, so do not expect meeting transcripts to become perfect.
There is also a licensing nuance worth checking: Apache 2.0 covers the models, but the training data is described as public audio plus synthetic data, and the model cards hold the details on provenance. For regulated industries, read them before deployment.
What to check before adopting
- Run your own audio. Collect 30 to 60 minutes of representative recordings, with transcripts, and compute WER for the 2B and Turbo CTC.
- Test your vocabulary. If you have product names or acronyms, use the 2,000-term keyword list and measure the difference.
- Measure throughput on your hardware. RTFx on the leaderboard reflects IBM's setup.
- Check language coverage. Languages outside the lists above are not claimed.
- Review the Apache 2.0 license and model cards for usage details and training-data notes.
What people are asking
Is the 2B better than Whisper? IBM compares against the leaderboard of open-weight models, not against a specific Whisper size in the post, so we cannot give a head-to-head from the source.
Can it run on a laptop? The 470M Turbo CTC is built for on-device work and has a browser demo; the 2B model has llama.cpp and Apple Silicon support, but memory needs depend on quantization.
Does keyword biasing work in every language? IBM says the 2,000-term lists apply in every supported language.
Related reading
- Granite Switch 4.1: chat, decisions and speech from one endpoint
- Tether TranscribePsy on-device speech recognition
- Cactus Whistle 16MB speech-to-text
- Gemini 3.5 Transcribe launch
- Microsoft MAI-Transcribe 2
- Hugging Face speech-to-speech voice agent guide
Figures are from IBM's October 21, 2026 post and its model cards; leaderboard positions change.