Ruby‑ASR refines the conventional ASR target into a span‑bound orthographic‑lexical‑reading sequence, addressing the fact that Japanese orthographic transcripts hide lexical‑reading distinctions. In standard systems the same written form receives a single label, so different pronunciations cannot be supervised and post‑hoc grapheme‑to‑phoneme conversion fails to recover them reliably. The ruby representation attaches the realized reading to each written span, enabling deterministic reconstruction of both the readable orthographic sentence and its lexical reading. We instantiate the approach on a Qwen3‑ASR backbone with two transcription conventions—subtitle‑style and verbatim‑style—and add a mora‑level CTC objective as auxiliary monotonic reading supervision. Experiments on five Japanese benchmarks show that the refined target improves lexical‑reading recovery without degrading orthographic readability. Checkpoints and inference code are released.
Review