Sanjid Hasan, Zaina Rahman, Suhita Islam Aurthi, MD Rashaduzzaman Rakib, Sadia Jahan Turabe, and Md. Wasif Rahman, "Beyond Recognition Accuracy: Efficient CPU Streaming ASR for Dysarthric Speech"
Abstract
Real-time dysarthric speech recognition is difficult because articulation varies widely across speakers while streaming systems must run under strict CPU latency limits. We present our Track 2 (Streaming ASR) system for the Speech Accessibility Project Challenge 2 (SAPC2). We fine-tuned a 66.1M-parameter causal Zipformer transducer on the SAPC2 training set, supplemented with TORGO recordings and 41,850 synthetic utterances from Qwen and Kokoro TTS. Training used randomized causal chunk sizes, SpecAugment, and acoustic perturbations. We also expanded the vocabulary from 500 to 510 tokens to emit digits directly while keeping the original token IDs intact. For CPU deployment, we exported the model to mixed-precision INT8 ONNX and streamed 100-ms audio chunks through sherpa-onnx. On the official Test1 leaderboard, our greedy model reached 22.18% CER and 31.73% WER—beating the 34.59% CER baseline—with a 1.38 s median stable first-token latency and 117 ms last-token delay. On a 501-utterance local set, 4-path beam search further lowered WER to 19.47% (12.59% CER). Perceptual annotations of over 437 samples show that recognition errors track perceived speech clarity (19.30% WER on clear vs. 57.28% on unclear speech) at least as strongly as medical etiology. Finally, we introduce Cadrate ASR, a 2.70M-parameter dual-rate CNN–GRU prototype that cuts CPU compute time below Zipformer, highlighting the trade-off between raw inference speed and recognition accuracy.