Ruby-ASR for Joint Orthographic and Lexical-Reading Recognition in Japanese
September 24, 2026
Ruby-ASR uses a Qwen3-ASR backbone with a mora-level CTC objective to solve the Japanese orthographic-lexical reading ambiguity. It employs a span-bound sequence that locally binds written spans to their realized pronunciations, enabling deterministic recovery of both views.
HOW THIS AFFECTS YOU
●
builderYou can build more accurate Japanese ASR systems that preserve both written text and correct phonetic readings.
●
researcherThe mora-level CTC objective provides a new way to supervise multi-view transcription tasks.