Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
Hugging Face
IBM updated three Granite Speech models for transcription, translation, and speaker tracking. One is more accurate, one is much faster, and one gets better at who said what.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
IBM’s Granite Speech line just got a broad refresh, and the pitch is simple: pick the model that fits the job instead of forcing one system to do everything badly. The newest release is granite-speech-5.0-2b, joined by updated versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus.
Granite Speech 5.0 2B is the headline model. It handles English, French, German, Spanish, Portuguese, and Japanese, and it translates to and from English in those languages, with extra support for English to Italian and Mandarin. IBM says it was trained on 390,000 hours of public audio, plus synthetic data for Japanese ASR, keyword biasing, and translation. The company is also leaning hard into control: language-ID prompting, keyword lists up to 2,000 terms, and optional punctuation and capitalization in the output.
On its own terms, the model is aimed at high-accuracy multilingual transcription and translation. IBM says it performs better on English short-form benchmarks and is more resilient to noise, reverberation, and distant microphones. It also does better on longer, more conversational translation, which is the part most demos conveniently ignore until real audio shows up.
Turbo CTC keeps the same 470M-parameter footprint but stops being English-only. The updated version adds multilingual transcription for English, Japanese, Spanish, German, French, Italian, and Portuguese, along with native punctuation and capitalization. IBM says it is now faster too: by compiling a single Conformer block and reusing its graph across the stack, warmup time fell and throughput reached 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at launch.
The 4.1 2B Plus update is for meetings, calls, interviews, and other multi-speaker recordings where timing matters as much as the transcript. IBM says it improves speaker attribution, timestamps, long-speech performance, and keyword biasing, while keeping the same architecture, prompt format, and API. In other words: a drop-in replacement, not a new migration project. IBM also says the models are under Apache 2.0 and are supported in Transformers, with additional support mentioned for vLLM, llama.cpp, and mlx-audio.
My take — AI-written commentary, not fact-checked reporting
The interesting bit here is not that IBM has three speech models. It’s that enterprise AI keeps splitting into narrow, practical tools instead of one grand model to rule them all. That’s boring in the best way: less hype, more transcripts that don’t embarrass anyone in a meeting.
Read more about this at: Hugging Face