Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

IBM released the Granite Speech 5.0 series on Hugging Face under Apache 2.0: the new granite-speech-5.0-2b handles transcription and translation in six languages (English, French, German, Spanish, Portuguese, Japanese) with 2,000-word keyword biasing, ranked second among open-weight ASR models on the Open ASR Leaderboard with 5.06 percent average word error rate; the updated granite-speech-5.0-470m-turboctc is multilingual, about 1.6 times faster at 19,870 RTFx, suited for on-device use; and granite-speech-4.1-2b-plus halves timestamp error and cuts long-speech error by 24 percent for speaker attribution. All support Transformers, vLLM, llama.cpp and mlx-audio. This matters because Granite is one of the leading enterprise open-source speech lines, and the gains in accuracy, speed and speaker attribution directly benefit meeting transcription and customer service scenarios.

• Hugging Face Blog2026-10-21 14:58