IBM releases Granite Speech 5.0 2B and updates Turbo CTC and 4.1 2B Plus speech models under Apache 2.0

IBM released granite-speech-5.0-2b, an open-weight speech model that handles transcription in six languages, translation to and from English, language-ID prompting and keyword lists of up to 2,000 terms. It was trained on 390,000 hours of public audio plus synthetic data. IBM also updated two existing models: the 470m TurboCTC model is now multilingual with punctuation and capitalisation and reaches 19,870 RTFx (a speed measure) on the Open ASR Leaderboard, up from about 12,600, and the 4.1-2b-plus weights are a drop-in replacement with long-form word error rate down from 13.53 to 10.22 and biased-word error rate down from 11.79 to 6.48. All three are Apache 2.0 and work with Transformers, vLLM, llama.cpp and mlx-audio.

Developers and enterprises who want self-hosted speech-to-text under a permissive licence get a credible, small option, and the llama.cpp and mlx-audio support means it can run on modest hardware. Teams already using the 4.1-2b-plus or TurboCTC models can swap in the new weights cheaply, and the gains in timestamp accuracy and keyword recognition matter most for meeting transcripts and call analysis. The impact is moderate: this is an incremental improvement in a crowded field, not a new capability, and two of the three items are updates rather than new models.

The claim that the three models together "cover the full range of enterprise speech work" is marketing framing. All benchmark figures come from IBM itself, and the #2 ranking on average English word error rate (5.06%) is explicitly unofficial, as of 15 October 2026, so it may not hold on the official leaderboard or on real noisy, accented or domain-specific audio. The speed jump is for a smaller CTC model and depends on hardware and batch settings, so it won't carry over to the 2B model.

1. 01Whether the 5.0-2b result is confirmed on the official Open ASR Leaderboard, and whether its English WER of 5.06% and #2 rank hold once independent submissions land.
2. 02Independent tests on non-English languages, noisy audio and long recordings, to check whether the stated gains in timestamp accuracy (AAS 210.86 to 98.54) and biased-word WER (11.79 to 6.48) reproduce outside IBM's own evaluation.
3. 03Adoption signals over the next one to two months: Hugging Face download counts, vLLM and llama.cpp integration issues, and any named enterprise deployments, compared with the 4.1 models.