IBM has released Granite Speech 5.0, an update to its open speech model family that adds one new model and revises two others, the company announced on October 21 through a Hugging Face post by its speech team. The three models split the work of enterprise speech into three jobs: high-accuracy transcription and translation, very fast transcription at the edge, and recordings where knowing who spoke, and when, matters as much as the words themselves.
All three models are released under the Apache 2.0 licence, with native support in Transformers and, for the new flagship, support in vLLM, llama.cpp and mlx-audio on Apple Silicon. IBM has published full evaluation results, training details and usage examples on each model card.
The New 2B Model Puts Accuracy And Translation First
The headline model, granite-speech-5.0-2b, succeeds granite-speech-4.1-2b at the same 2-billion-parameter size. It transcribes English, French, German, Spanish, Portuguese and Japanese, and translates to and from English in those languages, plus from English into Italian and Mandarin. IBM says it was trained on 390,000 hours of public audio, plus synthetic data built for Japanese speech recognition, keyword biasing and translation.
IBM reports stronger English accuracy on short-form benchmarks and better handling of noise, reverberation and distant-microphone speech than the model it replaces. By the company’s own count, the model ranks second among top open-weight speech recognition models on average English word error rate, at 5.06%, averaged over the Open ASR Leaderboard’s public English short-form test sets as of October 21. Those figures are IBM’s, reported on its own model cards.
The new model also adds controls that matter in production. Users can name the input language in the prompt to steer multilingual transcription, ask for punctuation and capitalisation in the output, and supply keyword lists of up to 2,000 terms to bias transcription and translation towards product names, people and acronyms in every supported language.
Turbo CTC Gets Multilingual, And About 1.6x Faster
The revised granite-speech-5.0-470m-turboctc keeps the 470-million-parameter footprint of the English-only encoder IBM introduced in August, and adds the three things the company says users asked for most. It now transcribes seven languages — English, Japanese, Spanish, German, French, Italian and Portuguese — produces punctuation and capitalisation natively, with no separate punctuation step, and was trained to be more robust in echoey rooms and far-field recordings.
It is also faster. By compiling a single Conformer block and reusing its graph across the stack, IBM cut warmup time and pushed the model to 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch — about 1.6 times faster at the same accuracy level, by IBM’s reporting. A WebGPU streaming demo runs the model in the browser, which is the use case the size is aimed at: transcription on device, without a round trip to a server.
The Plus Model Gets Better At Who Said What
The third update, granite-speech-4.1-2b-plus, keeps its name, architecture, prompt format and API, so existing users can swap it in as a drop-in replacement with improved weights. It is tuned for the recordings enterprises actually hold: meetings, support calls and interviews with several speakers running long.
IBM reports gains across all four areas it targeted. Word diarization error rate falls 7% and speaker-attributed word error rate falls 7% overall. Average timestamp shift is cut by more than half, from 210.86 to 98.54 across all speech, which tightens timing in captions, search and review. Word error rate on long-form test sets drops 24%, and errors on biased keywords fall 45% with 20-term keyword lists.
Releases like this one are best read against the wider open-model field, as our guide to how to read an AI model launch sets out, and against the standing argument over open-weight and closed AI models — Granite Speech sits firmly on the open side of that line, at sizes small enough to run outside a data centre.
Why It Matters
Granite Speech 5.0 matters because speech is where open models meet ordinary business data first. Meetings, support calls and interviews are recorded in volume, are often sensitive, and are exactly the workloads companies hesitate to send to a closed API. A 2-billion-parameter model that transcribes and translates in seven languages, a 470-million-parameter model fast enough to run in a browser, and a speaker-attribution model that drops into existing pipelines give those companies an on-premises option for each job, under a permissive licence. That also puts pressure on hosted speech services, including the voice platforms covered in our explainer on AI voices, cloning and text-to-speech: IBM is not selling generation here, it is giving away the listening half of the stack, and every workload that moves to an open model is one the paid APIs no longer bill for.
Background
IBM’s Granite family is the company’s open-model line, released under Apache 2.0 across language, code, time-series and speech models, and Granite Speech has been built up in public on Hugging Face. The speech line reached its 4.1 generation earlier in 2026, added the Turbo CTC speed model in August, and has now been refreshed as a three-model family in the 5.0 release.
Source: Hugging Face — IBM Granite Speech 5.0 announcement, AI Parabellum
0 Comments