Gradient-Based Speech-to-Text Alignment for Any ASR Model: From CTC to Speech LLMs

Jul 16, 2026

Upload by Asher

Speech-to-text alignment identifies each transcribed word’s start and end time in the audio. While CTC and transducer models provide alignment inherently and AED/LLM models typically derive word timing from attention weights, all signals are limited by the encoder frame grid. The document presents a generic gradient-based alignment for any differentiable ASR model: compute input gradients of teacher-forced token log probabilities, convert them to frame saliency, and recover word boundaries via dynamic programming. It needs no training or model changes, evaluates across sixteen models on read and spontaneous speech, and trades higher cost for usable timing, outperforming native alignment when the latter is weak.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction & Related Work

Background on alignment types across ASR families

Proposed gradient-based alignment method

Contributions and evaluation scope

Gradient-Based Speech-to-Text Alignment for Any ASR Model: From CTC to Speech LLMs Albert Zeyer , Ralf Schl¨uter , Hermann Ney Machine Learning and Human Language Technology Group, RWTH Aachen University, Aachen, Germany AppTek GmbH, Aachen, Germany {zeyer,schlueter,[ney](ney}@ml.rwth-aachen.de)[}](ney}@ml.rwth-aachen.de)[@ml.rwth-aachen.de](ney}@ml.rwth-aachen.de) arXiv :2607 .0683 1v 1 [ cs .CL] 7 Jul 2026 Abstract—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 encoderdecoders (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-framesaliency, 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 alignson 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. Index Terms—speech-to-text alignment, forced alignment, gradient-based attribution, automatic speech recognition, speech language models I. INTRODUCTION & RELATED WORK Speech-to-text alignment assigns each transcribed word a start and end time in the audio. It is classically produced by finding the best path through a Gaussian-mixture hidden Markov model (GM-HMM) which is still the most accurate aligner for read speech [1]–[5] . Current models, however, differ in what alignment they provide on their own [6]: connectionist temporal classification (CTC) [4],[7] and transducer models [8] have one by construction, whereas attention-based encoder-decoders (AED) [9]–[11] and speech large language models (LLMs) [12]–[14] do not, and their word timings are usually read off the attention weights instead. All of these signals live on the 20 to 80 ms encoder frame grid, which bounds their temporal precision. For AED, reading the timing off the cross-attention is established practice: Whisper emits coarse segment-level timestamp tokens [15], word-level timestamps follow from a dynamic time warping (DTW) of the decoder cross-attention [16] that fine-tuning can sharpen [17], and the alignment-bearing heads can even be selected by an unsupervised heuristic and refined with character-level teacher forcing without any training [18] . Speech LLM selfattention carries a similar alignment, used to constrain textto-speech synthesis [19] and as a policy for simultaneous translation [20]; forced alignment has separately been cast asa trained LLM task [21] . We study a generic gradient-based alignment that applies to any differentiable speech recognition model. For each transcript token we take the gradient of its teacher-forced log probability with respect to the input, reduce it to a per-framesaliency [22], and decode the resulting token-by-frame matrix into word boundaries with a single dynamic-programming pass (Fig. 1) . As it uses the gradient w.r.t. the input signal, it alignson the input grid rather than on the coarser encoder grid, and it can correct tempor

FAQs

What problem does the document address in speech-to-text systems?

It addresses speech-to-text alignment: determining temporal boundaries for each transcribed word within the audio.

How does the proposed method generate alignments for different ASR model families?

For each teacher-forced token, it takes the gradient of the token log probability with respect to the input, reduces it to per-frame saliency, and uses a single dynamic-programming pass to decode word boundaries.

What are the main trade-offs compared with native alignments?

The gradient-based alignment works across all model families and can be better when native forced alignment is weak, but it is usually more expensive due to one backward pass per token.

Jul 16, 2026

Upload by Asher