Granite Speech 4.1 2B
Granite Speech 4.1 2B remains a downloadable and hosted speech-language model. It supports recognition and bidirectional translation between English and French, German, Spanish, Portuguese or Japanese.
Model design and capabilities
Its dual-head CTC encoder has character and BPE outputs. Frame importance sampling focuses processing on informative audio. Prompt options control punctuation and capitalization, while keyword lists guide recognition. The 2B name describes the whole model, rather than only the language backbone.Granite Speech 4.1 2B model card
Practical use
Practical guidance. Keep this release as a comparison point when migrating an existing speech workflow to 5.0. Evaluate names, punctuation and translation direction with the exact same audio. Do not infer speaker labeling or word timestamps from the capabilities of the separate Plus variant.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable, Provider product |
| Model identifiers | ibm-granite/granite-speech-4.1-2b |
| License | Apache 2.0Granite Speech 4.1 2B model card |
| Training audio | 174,000 hours, with additional synthetic data.Granite Speech 4.1 2B model card |
| Ancestry | An intermediate Granite 4.0 1B Base checkpoint aligned to speech.Granite Speech 4.1 2B model card |
| Hosted access | watsonx: deploy on demand.IBM watsonx.ai foundation model library |
Related models
- Granite Speech 5.0 2B
IBM ยท Granite Speech
An audio-to-text model with a Granite 4.0 1B Base backbone for speech recognition in six languages.
Launch date Downloadable