IBM shared three updates to its Granite Speech collection: a new speech-language model, granite-speech-5.0-2b, and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. Together they cover the range of enterprise speech work: high-accuracy multilingual transcription and translation, extremely fast transcription at the edge, and multi-speaker recordings where speaker attribution and timestamps matter. All models are released under Apache 2.0, supported natively in Transformers, and accompanied by model cards that include evaluation results, architecture details, training data, and usage examples.

granite-speech-5.0-2b is the successor to granite-speech-4.1-2b at the same parameter count. It transcribes English, French, German, Spanish, Portuguese, and Japanese, and translates to and from English in those languages, plus English to Italian and Mandarin. It was trained on 390,000 hours of public audio, plus synthetic data built for Japanese ASR, keyword biasing, and translation. Compared with the 4.1 2B speech model, the new model delivers higher accuracy on English short-form benchmarks, better handling of noise, reverberation, and distant-microphone speech, language-ID prompting, better long-form and conversational translation, keyword lists up to 2,000 terms, and optional punctuation and capitalization on request. Capitalization F1 ranges from 89.7 to 99.5 across the reported test sets. On the Open ASR Leaderboard, it ranks #2 among top open-weight ASR models on average English WER at 5.06%; these results are labeled unofficial, as of 15 October 2026, and averaged over the leaderboard’s public English short-form test sets. The model ships with support in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.

The revised granite-speech-5.0-470m-turboctc keeps its 470M-parameter encoder footprint while adding three requested capabilities: multilingual transcription for English, Japanese, Spanish, German, French, Italian, and Portuguese; native punctuation and capitalization with no separate truecasing step; and training that explicitly targets reverberant audio for more robust far-field transcription. It is also faster. By compiling a single Conformer block and reusing its graph across the stack, IBM cut warmup time and pushed Turbo CTC to 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch, or about 1.6x faster at the same accuracy level.

The updated granite-speech-4.1-2b-plus is tuned for meetings, support calls, and long interviews with several speakers. It keeps the 4.1 name because the architecture, prompt format, and API are unchanged, making it a drop-in replacement with improved weights. This revision improves speaker attribution, with word diarization error rate down 7% and speaker-attributed WER down 7%. Average timestamp shift is cut by more than half for all speech, from 210.86 to 98.54, and improved for short speech from 29.27 to 26.20. Long-form WER drops 24%, from 13.53 to 10.22, and errors on biased keywords fall 45% with 20-term keyword lists, with biased-word WER moving from 11.79 to 6.48. All of these are error rates, so lower is better.

IBM’s selection guidance: use granite-speech-5.0-2b for high transcription accuracy, translation, or keyword biasing; use granite-speech-5.0-470m-turboctc for raw speed or on-device transcription, with a WebGPU streaming demo available; and use granite-speech-4.1-2b-plus for meetings and multi-speaker calls. The post also recommends pairing these models with Granite Guardian for production deployments that need additional risk detection.