AI Briefing
KOSign in

IBM releases Granite Speech 5.0 with improved ASR accuracy and 1.6x faster Turbo CTC

·2026.10.21 23:58

Key point

New 2B model ranks #2 on Open ASR Leaderboard while updated Turbo CTC achieves 19,870 RTFx.

1 / 2

Details

  • IBM releases three updates to the Granite Speech collection: new granite-speech-5.0-2b, updated granite-speech-5.0-470m-turboctc, and updated granite-speech-4.1-2b-plus.
  • granite-speech-5.0-2b achieves 5.06% average English WER, ranking #2 among top open-weight ASR models on the Open ASR Leaderboard.
  • granite-speech-5.0-470m-turboctc reaches 19,870 RTFx, a 1.6x speedup over its launch performance, while adding multilingual support and native punctuation.
  • granite-speech-4.1-2b-plus reduces word diarization error rate (WDER) by 7% and long-form WER by 24% for improved speaker attribution and timing.
  • All models are released under Apache 2.0 license with support in Transformers, vLLM, llama.cpp, and mlx-audio.
open-sourceasrspeech-recognitionibmgranite-speechhugging-face
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

HuggingFace BlogIBM Releases Granite Switch 4.1 Models Combining Chat, Decisions, and Speech on Single EndpointRedditAuroraASR-1M: A 949K-Parameter Speech Recognition Model Released Under Apache 2.0HuggingFace BlogNVIDIA Releases Nemotron 3 DiarizationRedditinclusionAI Releases Ming-3.0-audio-tiny, a 9.3B Parameter Native Audio-Language ModelKakaoKakao Releases Kanana-2-vision-encoder, Open-Sourcing Medium and Tiny WeightsGeekNewsArtCraft releases open-source IDE for AI image and video creation with 62-model catalog