IBM Granite Speech 5.0: Open Models for Transcription

IBM Granite Speech 5.0 is here, and it is a genuinely useful release for anyone building transcription into an app. Published on the Hugging Face blog on October 21, 2026, the update adds a new 2-billion-parameter speech-language model and refreshes two existing models, all under the permissive Apache 2.0 license. Together they cover high-accuracy transcription and translation, very fast on-device speech recognition, and multi-speaker meeting audio.

• New model: granite-speech-5.0-2b transcribes English, French, German, Spanish, Portuguese and Japanese, trained on 390,000 hours of public audio
• Accuracy: 5.06% average English word error rate, #2 among open-weight models on the Open ASR Leaderboard, per IBM's unofficial results as of October 15
• Speed: the 470M Turbo CTC model is now multilingual and about 1.6x faster at 19,870 RTFx
• Runtimes: Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon

What Is New in Granite Speech 5.0 2B?

The flagship granite-speech-5.0-2b succeeds the 4.1 2B model at the same size, which matters: you get better automatic speech recognition without a bigger memory footprint. IBM says English accuracy improves on short-form benchmarks and the model copes better with noise, echo and distant microphones.

Three features stand out for real deployments. Language-ID prompting lets you name the input language to steer transcription. Keyword biasing now accepts lists of up to 2,000 terms, so product names, people and acronyms come out spelled correctly in every supported language. And punctuation and capitalization can be requested in the prompt, with IBM reporting capitalization F1 scores between 89.7 and 99.5 across its test sets. The model also translates to and from English across its six languages, plus English to Italian and Mandarin.

How Fast Is the Granite Speech Turbo CTC Model?

When IBM introduced Turbo CTC in August, it was an English-only 470M-parameter encoder built for raw speed. The revision adds transcription in seven languages, native punctuation and capitalization, and better handling of reverberant rooms. IBM says compiling a single Conformer block and reusing its graph across the stack pushed throughput from roughly 12,600 RTFx to 19,870 RTFx on the Open ASR Leaderboard at the same accuracy level. For on-device or high-volume speech-to-text work, that is the one to try first, and IBM offers a WebGPU streaming demo that runs in the browser.

Better Speaker Attribution for Meetings

The updated granite-speech-4.1-2b-plus keeps its name because the architecture, prompt format and API are unchanged, making it a drop-in replacement with improved weights. IBM's own numbers show the average timestamp shift cut from 210.86 to 98.54, biased-keyword error rates falling from 11.79 to 6.48 with 20-term lists, and long-form word error rate dropping from 13.53 to 10.22. Word diarization error, which measures who said what, improves from 4.40 to 4.10.

Why Open Speech Models Matter

Speech recognition used to mean sending audio to a cloud API. Open-weight models like these let teams keep sensitive recordings, such as support calls or medical dictation, on their own hardware, and llama.cpp and mlx-audio support means a laptop can do the job. It builds on IBM's broader open push, including the Granite 4.2 open reasoning models, and joins a competitive field we tracked with Microsoft MAI Voice 2.1 streaming transcription. Note that all accuracy figures above are IBM's own and unofficial. For more model releases, see our AI coverage.

Sources: Hugging Face Blog (IBM Granite) — Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster — October 21, 2026; Hugging Face Open ASR Leaderboard — October 2026.