IBM Updates Granite Speech Models

IBM announced three updates to its Granite Speech collection on October 21, 2026: the new granite-speech-5.0-2b model and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The models target multilingual transcription and translation, fast edge or on-device transcription, and recordings with multiple speakers. All are released under the Apache 2.0 license and have native support in Transformers.

Granite Speech 5.0 2B, a successor to granite-speech-4.1-2b with the same parameter count, was trained on 390,000 hours of public audio plus synthetic data for Japanese speech recognition, keyword biasing, and translation. It transcribes English, French, German, Spanish, Portuguese, and Japanese; it translates to and from English in those languages, and from English to Italian and Mandarin. Users can name the input language in a prompt and provide keyword lists of up to 2,000 terms. Capitalization F1 scores range from 89.7 to 99.5 across IBM's test sets. IBM reports an average English word error rate (WER) of 5.06%, ranking the model #2 among top open-weight speech recognition models; the results are unofficial and use public English short-form test sets as of October 15, 2026.

The revised Turbo CTC retains its 470M-parameter size and adds transcription in Japanese, Spanish, German, French, Italian, and Portuguese alongside English. It now produces punctuation and capitalization natively and is more robust to reverberant and distant-microphone audio. IBM says it reaches 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at its August launch, or about 1.6x faster at the same accuracy. The speed increase came from compiling one Conformer block and reusing its graph across the model.

Granite Speech 4.1 2B Plus improves speaker attribution, timing, long-speech results, and keyword biasing while keeping the architecture, prompt format, and API unchanged. IBM reports that word diarization error rate and speaker-attributed WER each fell 7% overall; average timestamp shift fell from 210.86 to 98.54 for all speech and from 29.27 to 26.20 for short speech. WER on long-form test sets dropped 24%, from 13.53 to 10.22, while biased-word WER with 20-term lists fell 45%, from 11.79 to 6.48. The release also lists WDER moving from 4.40 to 4.10 and speaker-attributed cpWER from 30.50 to 28.48. IBM recommends Granite Guardian for production deployments that need additional risk detection.

IBM announced three updates to its Granite Speech collection on October 21, 2026: the new granite-speech-5.0-2b model and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The models target multilingual transcription and translation, fast edge or on-device transcription, and recordings with multiple speakers. All are released under the Apache 2.0 license and have native support in Transformers.

Granite Speech 5.0 2B, a successor to granite-speech-4.1-2b with the same parameter count, was trained on 390,000 hours of public audio plus synthetic data for Japanese speech recognition, keyword biasing, and translation. It transcribes English, French, German, Spanish, Portuguese, and Japanese; it translates to and from English in those languages, and from English to Italian and Mandarin. Users can name the input language in a prompt and provide keyword lists of up to 2,000 terms. Capitalization F1 scores range from 89.7 to 99.5 across IBM's test sets. IBM reports an average English word error rate (WER) of 5.06%, ranking the model #2 among top open-weight speech recognition models; the results are unofficial and use public English short-form test sets as of October 15, 2026.

The revised Turbo CTC retains its 470M-parameter size and adds transcription in Japanese, Spanish, German, French, Italian, and Portuguese alongside English. It now produces punctuation and capitalization natively and is more robust to reverberant and distant-microphone audio. IBM says it reaches 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at its August launch, or about 1.6x faster at the same accuracy. The speed increase came from compiling one Conformer block and reusing its graph across the model.

Granite Speech 4.1 2B Plus improves speaker attribution, timing, long-speech results, and keyword biasing while keeping the architecture, prompt format, and API unchanged. IBM reports that word diarization error rate and speaker-attributed WER each fell 7% overall; average timestamp shift fell from 210.86 to 98.54 for all speech and from 29.27 to 26.20 for short speech. WER on long-form test sets dropped 24%, from 13.53 to 10.22, while biased-word WER with 20-term lists fell 45%, from 11.79 to 6.48. The release also lists WDER moving from 4.40 to 4.10 and speaker-attributed cpWER from 30.50 to 28.48. IBM recommends Granite Guardian for production deployments that need additional risk detection.