Granite Speech Updates Announced
On October 21, 2026, IBM’s Granite team unveiled three significant updates to its Granite Speech collection: the introduction of the new speech-language model granite-speech-5.0-2b, along with revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. These updates aim to enhance high-accuracy transcription and translation, provide low-latency on-device transcription, and improve multi-speaker attribution.
granite-speech-5.0-2b: Enhanced Accuracy and Control
The granite-speech-5.0-2b model serves as the successor to granite-speech-4.1-2b, maintaining the same parameter count while introducing notable improvements. This model supports transcription in English, French, German, Spanish, Portuguese, and Japanese, and offers translation capabilities to and from English for these languages, as well as English to Italian and Mandarin. It was trained on approximately 390,000 hours of public audio, supplemented with synthetic data for Japanese ASR, keyword biasing, and translation.
Compared to its predecessor, the new model boasts:
- Improved English ASR Accuracy: Higher accuracy on short-form benchmarks and enhanced performance in noisy environments, reverberation, and distant microphone scenarios.
- Language-ID Prompting: Users can specify the input language in the prompt to guide multilingual transcription.
- Keyword Biasing: Supports lists of up to 2,000 terms to bias transcription and translation towards specific product names, people, and acronyms.
- Optional Punctuation and Capitalization: Users can request formatted output, with reported capitalization F1 scores ranging from 89.7 to 99.5 across test sets.
As of October 15, 2026, the model achieved an unofficial average English WER of 5.06%, ranking it #2 on the Open ASR Leaderboard.
Turbo CTC and 4.1-2b-plus Enhancements
The updated granite-speech-5.0-470m-turboctc model retains its 470M-parameter encoder while adding multilingual support for English, Japanese, Spanish, German, French, Italian, and Portuguese. It also features:
- Native Punctuation and Capitalization: Transcripts are generated ready to read, eliminating the need for separate punctuation or truecasing steps.
- Improved Robustness: Enhanced training for reverberant audio makes the model more effective in challenging acoustic environments.
- Increased Speed: By optimizing the model architecture, Turbo CTC’s warmup time has been reduced, achieving approximately 19,870 RTFx, a 1.6x speed increase from its initial launch.
The granite-speech-4.1-2b-plus model is designed as a drop-in replacement with several performance improvements, including:
- Speaker Attribution: More accurate identification of speakers, with a 7% reduction in word diarization error rate (WDER).
- Timestamps: Average timestamp shift (AAS) improved from 210.86 to 98.54, allowing for tighter timing in captions and reviews.
- Long-Speech Performance: WER on long-form test sets decreased by 24%.
- Keyword Biasing: Errors on biased keywords reduced by 45% with 20-term keyword lists.
Model Recommendations
IBM suggests the following model choices based on user needs:
- granite-speech-5.0-2b: Best for high transcription accuracy, translation, and keyword biasing.
- granite-speech-5.0-470m-turboctc: Ideal for raw speed and on-device transcription.
- granite-speech-4.1-2b-plus: Recommended for meetings and multi-speaker recordings.
All models are released under the Apache 2.0 license and support native integration in Transformers, with additional support for vLLM, llama.cpp, and mlx-audio for Apple Silicon.
Original source: Hugging Face Blog