We’re excited to share three updates to IBM’s Granite Speech assortment: a brand new speech-language mannequin, granite-speech-5.0-2b, and revised variations of two fashions the neighborhood already makes use of, granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. Collectively they cowl the complete vary of enterprise speech work: transcription and translation with excessive accuracy, extraordinarily quick transcription on the edge, and recordings the place who spoke, and when, matter as a lot as what was mentioned.
| Mannequin | Standing | What’s new | Finest for |
|---|---|---|---|
granite-speech-5.0-2b |
New | Extra correct, noise- and reverb-robust automated speech recognition (ASR), language-ID prompting, 2,000-term key phrase biasing, higher long-form translation | Excessive-accuracy multilingual transcription and translation |
granite-speech-5.0-470m-turboctc |
Up to date | Multilingual, native punctuation and capitalization, and extra strong to noise and reverb | Excessive-throughput and on-device transcription |
granite-speech-4.1-2b-plus |
Up to date | Improved speaker attribution, timestamps, key phrase biasing, and long-speech efficiency | Conferences, calls, and different multi-speaker recordings |
Granite Speech 5.0 2B: extra correct, extra strong, extra management
granite-speech-5.0-2b is the successor to granite-speech-4.1-2b on the similar parameter rely. It transcribes English, French, German, Spanish, Portuguese, and Japanese, and interprets to and from English in these languages, plus English to Italian and Mandarin. It was skilled on 390,000 hours of public audio, plus artificial information constructed for Japanese ASR, key phrase biasing, and translation.
What’s new in contrast with 4.1 2B Speech:
- Extra correct, extra strong English ASR: Greater accuracy on English short-form benchmarks, and higher dealing with of noise, reverberation, and distant-microphone speech.
- Language ID prompting: Identify the enter language within the immediate to steer multilingual transcription.
- Higher long-form and conversational translation: Translation holds up for longer with extra pure speech, not simply quick learn sentences.
- Key phrase lists as much as 2,000 phrases: Bias transcription and translation towards product names, individuals, and acronyms, in each supported language.
- Punctuation and capitalization on request: Ask for formatted output within the immediate; capitalization F1 ranges from 89.7 to 99.5 throughout our check units.
The mannequin ships below Apache 2.0 with assist in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
Chart 1: granite-speech-5.0-2b ranks #2 amongst prime open-weight ASR fashions on common English WER (5.06%). Decrease is healthier. Outcomes are official, as of 21 October 2026, and averaged over the Open ASR Leaderboard’s public English short-form check units.
Granite Speech 5.0 Turbo CTC: now multilingual, extra strong, and even sooner
Once we launched Turbo CTC in August, it was an English-only, 470M-parameter encoder constructed for pace. This revision retains that footprint and provides the three issues customers requested for many:
- Multilingual transcription: English, Japanese, Spanish, German, French, Italian, and Portuguese.
- Native punctuation and capitalization: Transcripts come out able to learn, with no separate punctuation or truecasing step.
- Extra strong in troublesome acoustics: Coaching that explicitly targets reverberant audio makes Turbo CTC extra strong in echoey rooms and far-field recordings.
It is also sooner. By compiling a single Conformer block and reusing its graph throughout the stack, we minimize warmup time and pushed Turbo CTC to 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch: about 1.6x sooner, on the similar accuracy stage.
Chart 2: Common WER vs. RTFx on the Open ASR Leaderboard (English short-form, 9 public check units, as of October 2026). Decrease WER and better RTFx are higher; RTFx is on a log scale.
Granite Speech 4.1 2B Plus: higher at who mentioned what, and when
The up to date granite-speech-4.1-2b-plus is tuned for the recordings enterprises even have: conferences, assist calls, and interviews with a number of audio system working lengthy. It retains the 4.1 identify as a result of the structure, immediate format, and API are unchanged, so it is a drop-in substitute for the present mannequin with improved weights. This revision improves 4 areas:
- Speaker attribution: Extra correct labeling of who’s talking. Phrase diarization error fee (WDER) drops 7%, and speaker-attributed WER (cpWER) drops 7% total.
- Timestamps: Common timestamp shift (AAS) is minimize by greater than half, yielding tighter timing in captions, search, and overview.
- Lengthy-speech efficiency: WER on long-form check units drops 24%.
- Key phrase biasing: Errors on biased key phrases fall 45% with 20-term key phrase lists.
| Functionality | Metric | Earlier launch | Up to date mannequin |
|---|---|---|---|
| Timestamps, all speech | AAS (common timestamp shift) | 210.86 | 98.54 |
| Timestamps, quick speech | AAS (common timestamp shift) | 29.27 | 26.20 |
| Key phrase biasing | Biased-word WER, 20-term lists | 11.79 | 6.48 |
| Lengthy-form speech | WER, long-form check units | 13.53 | 10.22 |
| Speaker attribution | WDER | 4.40 | 4.10 |
| Speaker-attributed transcripts | cpWER | 30.50 | 28.48 |
Desk 1: All metrics are error charges, so decrease is healthier. For extra detailed outcomes, see the mannequin card.
Get began
Undecided which mannequin to choose? Begin right here:
These fashions are all launched below an Apache 2.0 license with native assist in Transformers. Full analysis outcomes, structure particulars, coaching information, and utilization examples are on every mannequin card. We advocate pairing them with Granite Guardian for manufacturing deployments that want extra threat detection.
Give them a try to tell us what you suppose!
