Granite Speech 5.0 470M TurboCTC

Granite Speech 5.0 TurboCTC is a compact 470 million parameter automatic speech recognition (ASR) model for English only. IBM designed it for very high inference speed and deployment on laptops, smartphones and other edge devices.

The model uses a Conformer acoustic encoder with block self-attention, self-conditioning and temporal downsampling, ending in an output layer of 16,384 BPE units. The architecture has 16 Conformer blocks, reduces frame rates from 100 Hz to 12.5 Hz (8x subsampling) and applies self-conditioned CTC from the middle layer.

It was trained on roughly 60,000 hours of English audio from public corpora (including Common Voice 17, MLS, LibriSpeech, VoxPopuli, YODAS, AMI and Earnings-22) plus synthetic data, using a Connectionist Temporal Classification (CTC) objective. Inference is non-autoregressive with greedy decoding. The model is released under the Apache 2.0 license.