IBM Releases Granite Speech 5.0: Open-Source Voice Models Improve Accuracy and Speaker Attribution

IBM released the Granite Speech 5.0 series on October 21, 2026, comprising the new granite-speech-5.0-2b model and updated versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. This collection targets enterprise speech workloads, ranging from high-accuracy transcription and translation to ultra-fast edge processing and complex multi-speaker recordings. All models are released under the Apache 2.0 license with native support for frameworks like Transformers and vLLM.

The flagship granite-speech-5.0-2b succeeds the previous 4.1 version at the same parameter count. It demonstrates improved English short-form Automatic Speech Recognition (ASR) accuracy, with enhanced robustness against noise, reverberation, and distant-microphone speech. The model supports transcription and translation in English, French, German, Spanish, Portuguese, and Japanese, introducing language ID prompting to steer multilingual tasks via prompts. It also supports keyword biasing for up to 2,000 terms, improving recognition of product names, people, and acronyms. According to the Open ASR Leaderboard, it ranks second among top open-weight ASR models with an average English Word Error Rate (WER) of 5.06%.

For high-throughput and on-device deployment, the updated granite-speech-5.0-470m-turboctc retains its 470M parameter footprint while adding multilingual support (English, Japanese, Spanish, German, French, Italian, Portuguese) and native punctuation and capitalization. By compiling a single Conformer block and reusing its graph, inference speed increased to 19,870 RTFx (Real-Time Factor), approximately 1.6x faster than the initial launch rate of ~12,600 RTFx, maintaining the same accuracy level. This allows for higher concurrency or lower power consumption on equivalent hardware.