Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

What it claims

IBM released three Apache 2.0 speech models in October 2026, each aimed at a different job. granite-speech-5.0-2b targets accurate transcription, translation and keyword biasing. granite-speech-5.0-470m-turboctc targets speed and on-device use. granite-speech-4.1-2b-plus targets multi-speaker meetings and calls. The 2B model's 5.06% average English WER is IBM's own unofficial figure as of 15 October 2026. It is averaged over the Open ASR Leaderboard's public English short-form test sets.

How it works

granite-speech-5.0-2b replaces granite-speech-4.1-2b at the same parameter count. It was trained on 390,000 hours of public audio, plus synthetic data for Japanese ASR, 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 only into Italian and Mandarin. Several behaviours are set in the prompt:

• naming the input language to steer multilingual transcription;
• a keyword list of up to 2,000 terms, in every supported language;
• punctuated and capitalized output on request.

Turbo CTC keeps its 470M-parameter Conformer encoder. It was English-only at launch in August. It now transcribes English, Japanese, Spanish, German, French, Italian and Portuguese, so it covers Italian transcription, which the 2B model does not. It also adds native punctuation and capitalization, and training aimed at reverberant and far-field audio. The speedup comes from compiling a single Conformer block and reusing its graph across the stack, which also cuts warmup time.

granite-speech-4.1-2b-plus keeps the 4.1 architecture, prompt format and API. The release changes only its weights, so it is a drop-in replacement for the current model.

Evidence

• 2B: capitalization F1 ranges from 89.7 to 99.5 across IBM's test sets. Noise and distant-microphone gains over 4.1 2B are described but not given as numbers.
• Turbo CTC: 19,870 RTFx on the Open ASR Leaderboard, up from about 12,600 at launch (about 1.6x). IBM says accuracy is unchanged. The text gives no WER figure for it.
• 4.1 2B Plus: all figures are relative reductions against the previous 4.1 Plus weights.
• Word diarization error rate (WDER) falls 7%.
• Speaker-attributed WER (cpWER) falls 7%.
• Average timestamp shift falls by more than half.
• Long-form WER falls 24%.
• Errors on biased keywords fall 45% with 20-term keyword lists.

How to use it

All three models have native Transformers support. The 2B model also runs in vLLM, llama.cpp and mlx-audio on Apple Silicon. Turbo CTC has a WebGPU streaming demo that runs in the browser. The post gives no install or run commands and points to each model card for usage examples, architecture and training data. IBM recommends pairing the models with Granite Guardian in production deployments that need risk detection.

Limits

The leaderboard ranking and WER figures are self-reported and labelled unofficial. The 4.1 Plus gains are relative percentages with no absolute values in the text, and the full table is on the model card. The post does not say which runtimes besides Transformers support Turbo CTC or 4.1 Plus. Translation into Italian and Mandarin works from English only.