Search papers, labs, and topics across Lattice.
This paper introduces a gradient-based method for achieving speech-to-text alignment applicable to any differentiable automatic speech recognition (ASR) model, including attention-based encoder-decoders and speech large language models (LLMs). By taking the gradient of teacher-forced token log probabilities and decoding the resulting saliency matrix, the method provides temporal word boundaries directly on the input grid, bypassing the limitations of traditional alignment techniques. Evaluations across sixteen models demonstrate that while the gradient-based alignment is generally competitive with native methods, it excels in scenarios where native alignments are weak, such as with streaming models.
Gradient-based alignment can outperform traditional methods in challenging scenarios, offering a universal solution for precise speech-to-text mapping across diverse ASR models.
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 encoder-decoders (AED) and speech large language models (LLMs) do not, and their word timings are usually read off the attention weights instead. All of these signals live on the encoder frame grid, which bounds their temporal precision. We study a generic gradient-based alignment that applies to any differentiable ASR model. We take the gradient of each teacher-forced token log probability with respect to the input, reduce it to a per-frame saliency, and decode the resulting matrix into word boundaries with a single dynamic-programming pass. The method needs no training, no model modification and no alignment heads, works across all model families including the speech LLMs, and aligns on the input grid rather than on the coarser encoder grid. We evaluate it on sixteen models from four families, on read (TIMIT) and spontaneous (Buckeye) speech, each against the model's own native or attention-based alignment. We find that the gradient yields a usable alignment for every model, that it is usually somewhat behind a strong native aligner but better where the native alignment is weak, as for the streaming models, and that its main disadvantage is the cost of one backward pass per token.
