Granite Speech 5.0 470M TurboCTC
Granite Speech 5.0 470M TurboCTC is a non-autoregressive speech recognition model. Its current multilingual revision recognizes seven languages and produces capitalized, punctuated text.
Model design and capabilities
The encoder has 16 Conformer blocks with block attention, self-conditioning and temporal downsampling. A 16,384-unit multilingual BPE head produces CTC outputs. The current model uses 153,000 hours of public audio. It supplies transcription rather than general language reasoning or speech generation.Granite Speech 5.0 470M TurboCTC model card
Practical use
Practical guidance. TurboCTC is useful when transcription latency matters more than flexible language prompting. Measure word errors together with punctuation and names on your recording conditions. Keep downstream summarization as a separate step so that transcription errors remain visible.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-speech-5.0-470m-turboctc |
| License | Apache 2.0Granite Speech 5.0 470M TurboCTC model card |
| Parameters | 470MGranite Speech 5.0 470M TurboCTC model card |
| Languages | English, French, German, Spanish, Portuguese, Italian and Japanese.Granite Speech 5.0 470M TurboCTC model card |
Related models
- Granite Speech 5.0 2B
IBM · Granite Speech
An audio-to-text model with a Granite 4.0 1B Base backbone for speech recognition in six languages.
Launch date Downloadable - Granite Speech 5.0 470M TurboCTC NC
IBM · Granite Speech
A 470M English speech recognizer released under non-commercial, share-alike terms.
Launch date not confirmedResearch releaseDownloadable