IBM Granite Speech 5.0 adds multilingual ASR, faster Turbo CTC, improved diarization

IBM's Granite Speech 5.0 collection introduces granite-speech-5.0-2b, a new 2-billion-parameter model for multilingual transcription and translation across seven languages. The 470M-parameter Turbo CTC model gains multilingual support and reaches 19,870 RTFx on the Open ASR Leaderboard, up from ~12,600. An updated granite-speech-4.1-2b-plus improves speaker attribution accuracy by 7% and cuts timestamp error by more than half. All three models ship under Apache 2.0 with native support in Transformers, vLLM, llama.cpp, and mlx-audio.

What's new in each Granite Speech 5.0 model

• granite-speech-5.0-2b: New 2B-parameter model replacing 4.1 2B; supports English, French, German, Spanish, Portuguese, Japanese transcription and translation, plus English-to-Italian and English-to-Mandarin. Trained on 390,000 hours of public audio. Adds language-ID prompting, 2,000-term keyword biasing, and on-request punctuation/capitalization (F1 89.7–99.5). Ranks #2 on Open ASR Leaderboard English WER at 5.06%.
• granite-speech-5.0-470m-turboctc: Now multilingual (7 languages) with native punctuation/capitalization. Compiled Conformer block reuse pushes speed to 19,870 RTFx, ~1.6x faster than August 2026 launch, with improved robustness to reverberation.
• granite-speech-4.1-2b-plus: Drop-in weight update; WDER down 7%, cpWER down 7%, average timestamp shift cut from 210.86ms to 98.54ms, long-form WER down 24%, keyword biasing error down 45% with 20-term lists.

Why it matters

The Apache 2.0 licensing and broad inference framework support (Transformers, vLLM, llama.cpp, mlx-audio) positions these models as practical alternatives to proprietary speech APIs for enterprises and edge deployments. The Turbo CTC model's WebGPU streaming demo also lowers the barrier for browser-based transcription applications.