Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
Web page · https://huggingface.co/blog/ibm-granite/granite-speech-5
First seen by Oriel on .
AI summary
IBM introduces granite-speech-5.0-2b and updates two Granite Speech models for more accurate multilingual transcription, faster edge speech recognition, and improved speaker attribution.
- granite-speech-5.0-2b supports transcription in six languages, translation involving English, prompts for language identification, and keyword lists of up to 2,000 terms.
- Trained on 390,000 hours of public audio plus synthetic data, granite-speech-5.0-2b reports 5.06% average English WER and is released under Apache 2.0.
- The 470M-parameter granite-speech-5.0-470m-turboctc adds seven-language transcription and native punctuation; its reported speed rose to 19,870 RTFx, about 1.6 times its launch result.
- Updated granite-speech-4.1-2b-plus reduces long-form WER by 24%, timestamp shift by more than half, and errors on 20-term biased keywords by 45%.
- All three models support Transformers; the article recommends the 2B model for accuracy and translation, Turbo CTC for speed or on-device use, and 4.1-2b-plus for multi-speaker recordings.
#speech recognition #speech translation #speaker diarization #hugging face
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/ibm-granite/granite-speech-5 | https://huggingface.co/blog/ibm-granite/granite-speech-5 Read discussion |