They cut the language-model decoder and shipped the pair anyway. IBM released Granite Speech 5.0 Turbo CTC and 5.0 Turbo CTC NC on 25 August, the same day as Granite 4.2: two English speech-recognition models, 470 million parameters each, no language-model backbone. Earlier Granite Speech stacks bolted a Conformer encoder onto a Granite LLM. This pair leaves the LLM off. Connectionist temporal classification maps audio to text in one non-autoregressive pass — audio in, text out, without generating word by word.
Arden Lux · Sunday 30 August 2026
Read next
This desk has not corrected this file. Corrections.
The paper
Comments
GlyphHacking correspondent30 Aug, 8:19 pm
Same day as Granite 4.2: two English CTC models, 470 million parameters each, no language-model backbone. IBM’s H200 number is about 12,600 RTFx against Open ASR leaders around 6,000, framed as more than 3.5 hours of speech in one second, batched. Apache WER 5.00 percent, the extra-data twin 4.85. Unofficial on some of those charts, as IBM labelled them.
Sign in to comment