English word error is 5.06, on short clips

This IBM article on Hugging Face is dated Published October 21, 2026. The judgment first: the 5.06 percent for granite-speech-5.0-2b is the average word-error rate on the Open ASR leaderboard's public English short-form test sets. The page ranks it second among open-weight speech-recognition models and calls the figure official as of October 21. A lower word-error rate is better. This is not a score for long conversations, and it is not a translation score.

The 5.0 2B is the successor to the 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 recognition, keyword biasing, and translation. Against the 4.1 2B, the page lists higher robustness to noise, reverberation, and distant microphones, a prompt that can name the input language, keyword lists up to 2,000 terms, and punctuation on request. Capitalization F1 ranges from 89.7 to 99.5 across their test sets. Those are their test sets. This page does not name each set in the body.

The 470M Turbo CTC is a revision of the August release. The footprint stays the same. That earlier model was English only. This one adds transcription in English, Japanese, Spanish, German, French, Italian, and Portuguese, and it emits punctuation and capitalization directly. The page says training aimed at reverberant audio makes it more robust in echoey rooms and far-field recordings. The speed claim is that compiling one Conformer block and reusing its graph pushed Turbo CTC to 19,870 RTFx on the Open ASR leaderboard, up from roughly 12,600 RTFx at launch, about 1.6 times faster, at the same accuracy level. The 1.6 times and the same accuracy are the page's statement. The chart note also says English short-form, nine public test sets, RTFx on a log scale, as of October 2026.

granite-speech-4.1-2b-plus keeps the 4.1 name because the architecture, prompt format, and API are unchanged. The weights are what changed. The figures on the page are error rates, so lower is better. Average timestamp shift on all speech falls from 210.86 to 98.54. On short speech it falls from 29.27 to 26.20. With 20-term keyword lists, biased-word error falls from 11.79 to 6.48. Word error on long-form test sets falls from 13.53 to 10.22. Speaker-label error, WDER, falls from 4.40 to 4.10, and speaker-attributed cpWER falls from 30.50 to 28.48. The body's approximations, about 7 percent, more than halving the timestamp shift, a 24 percent drop in long-form word error, and a 45 percent drop in keyword error, summarize that table. The finer list of test sets is on the model card, not in the body of this page.

All three models are released under Apache 2.0, with native support in Transformers. The 2B page also names vLLM, llama.cpp, and mlx-audio for Apple Silicon. The page recommends pairing them with Granite Guardian for production deployments that need extra risk detection. That is a recommendation, not a comparison experiment on this page. The three-line picker is also the page's own guidance: use the 5.0 2B for accuracy, translation, or keyword biasing; use Turbo CTC for speed or on-device transcription; use the 4.1 2B Plus for meetings and multi-speaker calls. Guidance is not a test result.