Junlong Shen, "Coordinating Offline and Streaming Speech Recognizers for SAPC2"
Abstract
We describe team JLShen's offline and streaming systems for the Second Speech Accessibility Project Challenge (SAPC2). Both combine complementary recognizers, but the resource and latency constraints change how those recognizers interact. The offline system fuses token-and-duration transducer posteriors during search, re-ranks hypotheses with speech models, and adapts its encoder using confident earlier outputs from the same speaker. The streaming system pairs a cache-aware CTC partial recognizer with a quantized final transducer. Its central design issue is first-word agreement: a final correction can undo the latency benefit of an early partial transcript. We address this through emission training, speculative endpoint decoding, and policies that constrain first-word revisions. On the public evaluation set, the reported offline system achieves 8.0650% character error rate (CER). Streaming configurations span 8.8455% CER at 613.8 ms total latency and 11.9272% at 400.8 ms. These are measured operating points of complete systems; the comparisons do not isolate individual components' effects.