Researchers present a generic gradient-based method to find temporal word boundaries in audio for any differentiable Automatic Speech Recognition (ASR) model, including those that do not natively provide alignment like attention-based encoder-decoders and speech LLMs. The approach computes the gradient of each teacher-forced token log probability with respect to the input, reduces it to per-frame saliency, and decodes word boundaries via dynamic programming.
- The method requires no training, model modification, or additional alignment heads.
- It aligns on the finer input grid rather than the coarser encoder frame grid used by CTC or transducer models.
- Evaluation across sixteen models from four families shows it yields usable alignments for every model tested.
- Performance is generally slightly behind strong native aligners but superior where native alignment is weak, such as in streaming models.
The primary disadvantage identified is the computational cost of performing one backward pass per token.