Skip to content
Neural Daily

ModelsNews

IBM Granite Speech 5.0: sharper ASR, faster Turbo CTC

IBM updated three models in its Granite Speech family. The new granite-speech-5.0-2b is more accurate and noise-robust, Turbo CTC went multilingual, and 4.1 2B Plus is better at tracking who said what and when.

By: Mikola Máté Patrik · · 3 min read

IBM Granite Speech cover image: the Granite Speech wordmark, a sound wave and a cartoon rock wearing a headset
Image: IBM · Press kit · Official blog social image (og:image) · Source

What changed?

On October 21, IBM announced three models in its Granite Speech collection on the Hugging Face blog (blog post). One is brand new; the other two are updates to models the community already uses. All three are Apache 2.0 licensed with native Transformers support.

granite-speech-5.0-2b: the new flagship

The new model succeeds granite-speech-4.1-2b at the same parameter count. It transcribes English, French, German, Spanish, Portuguese and Japanese, translates between those languages and English, and also from English to Italian and Mandarin. It was trained on 390,000 hours of public audio plus synthetic data.

  • More accurate English ASR that handles noise, reverberation and distant microphones better.
  • The input language can be named in the prompt to steer multilingual transcription.
  • Keyword lists of up to 2,000 terms bias recognition toward product names, people and acronyms.
  • Punctuation and capitalization on request, with capitalization F1 between 89.7 and 99.5 on IBM's test sets.
  • Supported in Transformers, vLLM, llama.cpp and mlx-audio on Apple Silicon.

By IBM's own unofficial measurement, the model ranks second among open-weight ASR models on the Open ASR Leaderboard's English short-form sets, with an average WER (word error rate) of 5.06%. First place is Qwen3-ASR-1.7B at 4.94%.

IBM chart: English short-form WER of granite-speech-5.0-2b versus top open-weight ASR models
IBM · Press kit · Source

Source: IBM, Hugging Face blog (unofficial results as of October 15, 2026)

Turbo CTC: multilingual and 1.6x faster

granite-speech-5.0-470m-turboctc was English-only when it launched in August. It now transcribes English, Japanese, Spanish, German, French, Italian and Portuguese with native punctuation. IBM compiles a single Conformer block and reuses its graph across the stack, lifting the model to 19,870 RTFx on the Open ASR Leaderboard from roughly 12,600, at the same accuracy. A WebGPU streaming demo runs it in the browser.

4.1 2B Plus: better at who said what

granite-speech-4.1-2b-plus keeps its architecture, prompt format and API, so the new weights are a drop-in replacement. IBM's figures show word diarization error rate (WDER) falling from 4.40 to 4.10 and speaker-attributed cpWER from 30.50 to 28.48. Average timestamp shift dropped from 210.86 to 98.54, long-form WER from 13.53 to 10.22, and biased-keyword WER with 20-term lists from 11.79 to 6.48.

Why it matters

If you build meeting transcription, call analytics or captioning, you now get near-top English accuracy, keyword biasing and translation from a permissively licensed 2B model. Turbo CTC's speed makes it a strong candidate for on-device and high-throughput pipelines.

What to watch

IBM ran the benchmarks itself and the leaderboard ranking is unofficial. Test on your own audio, especially for languages and acoustic conditions outside the published sets.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles