Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
IBM updates its Granite Speech collection with three new and revised speech-language models, enhancing transcription, translation, and speaker attribution across multiple languages and use cases.
IBM has introduced granite-speech-5.0-2b, a new 2-billion-parameter speech-language model designed for high-accuracy transcription and translation in English, French, German, Spanish, Portuguese, Japanese, Italian, and Mandarin. The model builds on its predecessor with expanded language support and was trained on 390,000 hours of public audio plus synthetic data tailored for Japanese ASR, keyword biasing, and translation tasks. It is released under the Apache 2.0 license with compatibility across Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
The revised granite-speech-5.0-470m-turboctc model, previously English-only, now includes multilingual support while maintaining its 470-million-parameter encoder optimized for speed. By restructuring its Conformer block and reusing computational graphs, the model achieves 19,870 RTFx on the Open ASR Leaderboard, a 1.6x speed improvement over its launch performance without sacrificing accuracy.
Granite-speech-4.1-2b-plus has been updated to better handle enterprise recordings such as meetings, support calls, and multi-speaker interviews. The revision improves error rates in diarization, punctuation, capitalization, and speaker change detection while retaining the same architecture, prompt format, and API for seamless integration as a drop-in replacement.
All three models are available under the Apache 2.0 license with native support in the Transformers library. IBM recommends pairing them with Granite Guardian for production deployments requiring additional risk detection, and invites users to test the models and provide feedback.