IBM's Granite Speech collection is getting a coordinated update, according to a Hugging Face post from the IBM Granite team. The release includes a new flagship model, granite-speech-5.0-2b, plus revised versions of the edge-oriented granite-speech-5.0-470m-turboctc and the multi-speaker granite-speech-4.1-2b-plus. The three models are positioned for different enterprise workloads: high-accuracy transcription and translation, fast on-device transcription, and recordings where speaker attribution matters.
The new 2B model is a successor to granite-speech-4.1-2b at the same parameter count. It transcribes English, French, German, Spanish, Portuguese, and Japanese, and translates between English and those languages, plus English to Italian and Mandarin. IBM says it was trained on 390,000 hours of public audio plus synthetic data, and its improvements include language-ID prompting, punctuation and capitalization on request, and keyword biasing lists of up to 2,000 terms. The model ships under Apache 2.0 with support in Transformers, vLLM, llama.cpp, and mlx-audio. On the Open ASR Leaderboard's public English short-form sets, the post reports an average WER of 5.06%, ranking #2 among top open-weight ASR models as of 15 October 2026.
The updated Turbo CTC keeps its 470M-parameter footprint but expands from English-only to seven languages, adds native punctuation and capitalization, and targets reverberant and far-field audio. The authors report it reached 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch, about 1.6x faster at the same accuracy level. The revised 4.1-2b-plus is a drop-in replacement with unchanged architecture, prompt format, and API; the post reports a 7% drop in word diarization error rate and speaker-attributed WER, a more than halving of average timestamp shift, a 24% drop in long-form WER, and 45% fewer errors on biased keywords with 20-term lists.
Because this is a single vendor announcement, the performance claims are not independently verified. The post's selection guidance is straightforward: choose the 2B model for accuracy, translation, and large keyword lists; Turbo CTC for high-throughput or on-device work; and 4.1-2b-plus for meetings, calls, and other multi-speaker recordings.