IBM has announced three significant updates to its Granite Speech model collection, introducing the new granite-speech-5.0-2b and revised versions of the existing Turbo CTC and 4.1 2B Plus models. The new lineup covers a broad spectrum of enterprise needs, from high-accuracy transcription to ultra-fast edge processing.
Granite Speech 5.0 2B: Precision and Control
The granite-speech-5.0-2b model succeeds the 4.1 version, maintaining the same parameter count while offering enhanced robustness. It supports transcription in six languages (English, French, German, Spanish, Portuguese, and Japanese) and translates to and from English, with additional support for Italian and Mandarin. According to IBM, the model ranks second among top open-weight ASR models, boasting an average English Word Error Rate (WER) of 5.06% on public benchmarks.
- Language ID Prompting: Allows users to specify the input language to guide multilingual transcription.
- Keyword Biasing: Supports lists of up to 2,000 terms to improve the recognition of product names and acronyms.
- Formatted Output: Punctuation and capitalization can be requested directly within the prompt, with high F1 scores across test sets.
Turbo CTC and 4.1 2B Plus: Speed and Meetings
The updated Granite Speech 5.0 Turbo CTC focuses on raw speed, reaching 19,870 RTFx—a 1.6x increase over its launch performance. It is now multilingual and more resilient in challenging acoustic environments, such as reverberant rooms and far-field recordings.
For meetings and interviews, the Granite Speech 4.1 2B Plus provides improved speaker attribution. The Word Diarization Error Rate (WDER) has dropped by 7%, while performance on long-form speech saw a 24% reduction in WER. This model serves as a drop-in replacement for the previous 4.1 version, maintaining the same architecture and API.
All models are released under the Apache 2.0 license, ensuring broad accessibility for the developer community and native support in frameworks like Transformers and vLLM.
