IBM Granite Speech 5.0 Boosts Multilingual ASR Speed and Accuracy
IBM's Granite Speech 5.0 adds multilingual ASR, translation and faster edge transcription, while improving speaker labels, timestamps and keyword biasing.
Summary
On October 21, 2026, IBM introduced granite-speech-5.0-2b and updated granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The new 2B model transcribes English, French, German, Spanish, Portuguese and Japanese, translates between English and the other five languages, and translates English into Italian and Mandarin. Trained on 390,000 hours of public audio plus synthetic Japanese ASR, translation and keyword data, it adds language prompts, long-form conversational translation, 2,000-term biasing and requested formatting, with capitalization F1 of 89.7 to 99.5. Its unofficial 5.06% average English WER ranked second among open-weight ASR models on October 15, 2026.
The updated 470M Turbo CTC adds seven-language transcription, native punctuation and capitalization, and greater noise and reverberation tolerance. Reusing one compiled Conformer block lifts Open ASR Leaderboard speed from roughly 12,600 RTFx at its August launch to 19,870 RTFx, about 1.6 times faster at unchanged accuracy, based on nine English short-form tests in October 2026.
The drop-in 4.1 2B Plus update improves meetings, calls and interviews without changing its architecture, prompts or API. All-speech timestamp shift falls from 210.86 to 98.54, short-speech shift from 29.27 to 26.20, biased-word WER from 11.79 to 6.48, long-form WER from 13.53 to 10.22, WDER from 4.40 to 4.10 and cpWER from 30.50 to 28.48. All three models use Apache 2.0 and native Transformers support; the new 2B also supports vLLM, llama.cpp and mlx-audio. IBM offers a WebGPU Turbo demo, full model cards and recommends Granite Guardian for added production risk detection.
Positives
- Granite Speech 5.0 2B ranked second among open-weight ASR models with an unofficial 5.06% average English WER.
- Turbo CTC reached 19,870 RTFx, about 1.6 times its roughly 12,600 RTFx August launch performance without reducing accuracy.
- Granite Speech 4.1 2B Plus cut long-form WER 24% and errors on 20-term biased keyword lists 45%.
- Keyword lists of up to 2,000 terms can prioritize product names, people and acronyms across every supported 2B language.
- All three models carry Apache 2.0 licenses and native Transformers support, improving deployment flexibility for enterprises and developers.
Risks & concerns
- The 5.06% English WER ranking was unofficial as of October 15, 2026, and placed Granite Speech 5.0 2B second rather than first.
- Turbo CTC benchmarking covered nine English short-form test sets, leaving multilingual and long-form performance unmeasured by that comparison.
- Granite Speech 4.1 2B Plus still records 28.48 cpWER and 10.22 long-form WER after its improvements.
- Italian and Mandarin are English translation targets for the 2B model, but IBM does not list them as transcription languages.
- IBM recommends pairing the speech models with separate Granite Guardian software when production deployments require additional risk detection.