ZeroHour
arXiv cs.AI / cs.LG / cs.CLpublished ()ingested Brian Kingsbury

Design of the IBM Granite 5.0 TurboCTC ASR Model

infoModel releaseimportance 32
AI summary · glm-5.3-flash

IBM details Granite 5.0 TurboCTC, a 470M-parameter Conformer-based English ASR model on the Open ASR speed-accuracy Pareto frontier, released on Hugging Face.

IBM describes Granite 5.0 Turbo CTC, a 470 million parameter encoder-only ASR model built on Conformer blocks with pyramidal temporal subsampling, strided depthwise convolutions, chunk-wise self-attention, and conditioning on middle-layer predictions. Training uses only publicly available data, a novel Muon optimizer, and balanced data sampling. Inference optimizations make it twice as fast as the fastest competitor while sitting on the Open ASR leaderboard Pareto frontier for English short-form ASR. Weights are released on Hugging Face under a permissive license.

  • 470M-parameter encoder-only Conformer model with strided depthwise convolutions and chunk-wise self-attention
  • Trained only on publicly available data using a novel Muon optimizer and balanced sampling
  • Twice as fast as the fastest competitor on the Open ASR leaderboard Pareto frontier
  • Released on Hugging Face (ibm-granite/granite-speech-5.0-470m-turboctc) under a permissive license
OrganizationsIBM
Full article135 words · extracted from arxiv.org · click to collapse

We describe the architecture, training methodology and inference speedups of Granite 5.0 Turbo CTC, a 470 million parameter encoder-only model with an excellent speed-accuracy tradeoff. The architecture uses pyramidal temporal subsampling within Conformer blocks using strided depthwise convolutions, block-diagonal (chunk-wise) self-attention, and conditioning on intermediate predictions from the middle layer. Training highlights are the use of only publicly available data, the novel use of a Muon optimizer, and balanced data sampling. Inference speedups include replacing 1 x 1 convolutions with linear layers and optimizing the attention computation in the Conformer blocks. Collectively, these result in a model that is on the speed-accuracy Pareto frontier of the Open ASR leaderboard for English short-form ASR while being twice as fast as the fastest competitor. The model can be used under a permissive license and downloaded from https://huggingface.co/ibm-granite/granite-speech-5.0-470m-turboctc.

Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.20104