IBM adds language and keyword controls to Granite Speech
The new 5.0 2B model targets multilingual transcription and translation, with prompts that steer recognition toward the input language and relevant terms.
IBM’s October 21 announcement introduces Granite Speech 5.0 2B for turning speech into text or translating it. Developers can name the input language and supply up to 2,000 keywords to steer recognition toward names and acronyms.
The 2B model documentation lists transcription in English, French, German, Spanish, Portuguese and Japanese. For best results, IBM recommends audio segments of up to about two minutes and splitting longer recordings before decoding. That is a recommended processing workflow, not a claim that every longer recording fails. The documentation provides usage examples for Transformers, vLLM, llama.cpp and mlx-audio. It also warns that responses can be inaccurate, biased or unwanted.
The companion Turbo CTC model addresses transcription with low delay or high throughput. Its October 21 update expands the previously English-only model to English, French, German, Spanish, Portuguese, Italian and Japanese. Transcripts include punctuation and capitalization. This is a 470-million-parameter model, and its documented native Transformers integration requires version 5.16.0 or later.
For recordings where identifying speakers and their timing matters, IBM’s release guide points to the separately updated 4.1 2B Plus.
The October 22 MarkTechPost report also covers the release and directs readers to IBM’s blog and model documentation.
Sources(4)
- IBM releaseChecked:
- 2B modelChecked:
- Turbo modelChecked:
- MarkTechPost reportChecked:
