arxivcs.CLcs.AIcs.CVcs.LG2026-07-07
Gradient-Based Speech-to-Text Alignment for Any ASR Model: From CTC to Speech LLMs
Albert Zeyer, Ralf Schlüter, Hermann Ney
Speech-to-text alignment means finding the temporal boundaries of each word in the audio. Some models provide such an alignment directly and others do not. Connectionist temporal classification (CTC) and transducer models have an alignment by construction, whereas attention-based…