MDL-1669EST.2026 · IDX.732
AudioIn production

IBM Granite Speech 5.0 470M TurboCTC

IBM · USA · 2026

IBM opened its 5.0 generation not with a chatbot but with a 470-million-parameter transcriber small enough for a phone - and published it twice, under two different licences.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Granite Speech 5.0 470M TurboCTC, published on 25 August 2026, is the first model IBM has released under the Granite 5.0 name, and it is worth noting what kind of model that is. Not a flagship language model, not a reasoning system - a 470-million-parameter English transcriber designed to run on laptops, phones and edge devices. It drops the language model entirely: where Granite Speech 4.1 bolts a speech encoder onto a Granite LLM, this one is a bare conformer encoder that emits text directly, without generating token by token. That is what makes it fast. The architecture is built around throughput. Sixteen conformer blocks, block attention over windows of 128 frames, and an aggressive eightfold time compression that takes the audio stream from 100 to 12.5 frames per second before the network does its main work. Decoding is non-autoregressive and greedy, so there is no beam search and no sequential generation - the whole utterance is read out in one pass. Training took ten days on eight H100 cards, which by the standards of this field is close to nothing. The most instructive thing about this release is not in the model at all but in how it was published. IBM issued the same design twice on the same day. This version, trained on 60,000 hours, carries a plain Apache 2.0 licence and may be used commercially. A twin, granite-speech-5.0-470m-turboctc-nc, was trained on 75,000 hours - the extra 14,900 coming from GigaSpeech and SPGI Speech - and carries CC BY-NC-SA 4.0, which forbids commercial use. The licence of a speech model is set by the licence of the recordings it heard, and IBM chose to make that trade visible rather than quietly restrict everything. For a Polish reader the limit is simple: this model is English-only, with no multilingual claim. For anyone needing other languages, Granite Speech 4.1 2B remains the family's multilingual option.

#open-weights#speech recognition#transcription#on-device#apache-2.0#IBM
Official website

News

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review