Omni-temporal Classification (OTC) improves ASR performance by adding wildcard paths to the CTC alignment graph at the token level. This allows models to bypass transcription ambiguities like spelling or pronunciation variations without losing supervision for the rest of the word.
HOW THIS AFFECTS YOU
●
builderThis improves the reliability of speech-to-text pipelines when dealing with non-verbatim or accented audio.
●
researcherYou can implement token-granularity wildcard arcs to make ASR training more robust to noisy reference transcripts.