We’re excited to share three updates to IBM’s Granite Speech collection: a new speech-language model, granite-speech-5.0-2b, and revised versions of two models the community already uses, granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. Together they cover the full range of enterprise speech work: transcription and translation with high accuracy, extremely fast transcription at the edge, and recordings where who spoke, and when, matter as much as what was said.
| Model | Status | What’s new | Best for |
|---|---|---|---|
granite-speech-5.0-2b |
New | More accurate, noise- and reverb-robust automatic speech recognition (ASR), language-ID prompting, 2,000-term keyword biasing, better long-form translation | High-accuracy multilingual transcription and translation |
granite-speech-5.0-470m-turboctc |
Updated | Multilingual, native punctuation and capitalization, and more robust to noise and reverb | High-throughput and on-device transcription |
granite-speech-4.1-2b-plus |
Updated | Improved speaker attribution, timestamps, keyword biasing, and long-speech performance | Meetings, calls, and other multi-speaker recordings |
Granite Speech 5.0 2B: More Accurate, More Robust, More Control
granite-speech-5.0-2b is the successor to granite-speech-4.1-2b at the same parameter count. It transcribes in multiple languages—English, French, German, Spanish, Portuguese, and Japanese—and facilitates translations between English and the mentioned languages, plus Italian and Mandarin. Trained on a colossal dataset of 390,000 hours of public audio and synthetic data crafted specifically for Japanese ASR, this model sets a new standard in the realm of speech recognition and translation.
What’s new compared to the previous version? Here’s a quick rundown:
- More Accurate, More Robust English ASR: Enhanced accuracy on English short-form benchmarks and improved handling of noise, reverberation, and distant-microphone speech.
- Language ID Prompting: Introducing language identification in the input prompt allows more accurate multilingual transcription.
- Better Long-Form and Conversational Translation: Translation quality remains high for longer speeches, yielding a more natural flow than simply short read sentences.
- Keyword Lists Up to 2,000 Terms: Users can bias transcription towards specific terms, including product names and acronyms, across supported languages.
- Punctuation and Capitalization on Request: Request formatted output in the prompt, achieving capitalization F1 scores ranging from 89.7 to 99.5 across test sets.
This state-of-the-art model is released under Apache 2.0 with ready support in popular frameworks like Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
Chart 1: 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.
Granite Speech 5.0 Turbo CTC: Now Multilingual, More Robust, and Even Faster
Initially launched as an English-only, 470M-parameter encoder optimized for speed, the granite-speech-5.0-470m-turboctc model has undergone significant upgrades. Users requested enhancements, and we delivered:
- Multilingual Transcription: Support for English, Japanese, Spanish, German, French, Italian, and Portuguese.
- Native Punctuation and Capitalization: Transcriptions are now ready to read out of the box, eliminating the need for post-processing punctuation.
- More Robust in Difficult Acoustics: Training now explicitly targets handling reverberant audio, enhancing performance in echoey environments and far-field recordings.
Speed improvements are notable. By compiling a single Conformer block and reusing its graph across the stack, we achieved a performance boost, pushing Turbo CTC to 19,870 RTFx on the Open ASR Leaderboard, markedly improving from roughly 12,600 RTFx at launch—about 1.6 times faster while maintaining accuracy.
Chart 2: Average WER vs. RTFx on the Open ASR Leaderboard (English short-form, 9 public test sets, as of October 2026). Lower WER and higher RTFx are better; RTFx is measured on a logarithmic scale.
Granite Speech 4.1 2B Plus: Better at Who Said What, and When
The updated granite-speech-4.1-2b-plus model is tailored for enterprise recordings, such as meetings and support calls featuring multiple speakers. Though it retains the 4.1 name—keeping the architecture, prompt format, and API unchanged—it boasts enhanced capabilities across several important facets:
- Speaker Attribution: Markedly improved labeling and identification of speakers, with a 7% reduction in word diarization error rate (WDER) and a corresponding 7% improvement in speaker-attributed word error rate (cpWER).
- Timestamps: Average timestamp shifts have halved, allowing for more accurate captioning, searching, and review functionalities.
- Long-Speech Performance: Word error rate (WER) on long-form tests improved by an impressive 24%.
- Keyword Biasing: Accuracy on biased keywords is enhanced, with a 45% improvement on 20-term keyword lists.
| Capability | Metric | Previous Release | Updated Model |
|---|---|---|---|
| Timestamps, all speech | AAS (average timestamp shift) | 210.86 | 98.54 |
| Timestamps, short speech | AAS (average timestamp shift) | 29.27 | 26.20 |
| Keyword biasing | Biased-word WER, 20-term lists | 11.79 | 6.48 |
| Long-form speech | WER, long-form test sets | 13.53 | 10.22 |
| Speaker attribution | WDER | 4.40 | 4.10 |
| Speaker-attributed transcripts | cpWER | 30.50 | 28.48 |
Table 1: All metrics are error rates, so lower is better. For more detailed information, please refer to the model card.
Get Started
Not sure which model to pick? Start here:
All models are released under an Apache 2.0 license with native support in Transformers. For comprehensive evaluation results, architecture details, training data, and usage examples, check the individual model cards. We recommend pairing these models with Granite Guardian for production deployments needing additional risk detection.
Give them a try and let us know what you think!
Inspired by: Source
