arXiv · 2611.18667
Speaking-Clock Mamba: Phoneme-Conditioned State-Space Modeling for Visual Forced Alignment
Abstract
Visual forced alignment (VFA) locates phoneme and word boundaries in a silent talking-face video given its transcript. Existing methods mainly enforce transcript order at the decoding stage, leaving the encoder without an explicit mechanism to anchor its temporally evolving representation to transcript progress. As a result, phoneme-specific evidence is blurred and alignment drift accumulates over long sequences. In this paper, we propose Speaking-Clock Mamba, which makes transcript progress part of the temporal recurrence. A selective state-space (Mamba) backbone is inserted between the visual frontend and the Conformer cross-attention encoder, and a Phoneme-Conditioned Selective Scan (PCSS) conditions its input-dependent step size on a supervised, structurally monotone speaking clock that estimates progress through the transcript. Since causal and bidirectional variants of this backbone make complementary errors, a training-free selector picks the more confident alignment per utterance. Experiments on LRS2 and LRS3 show consistent gains over protocol-matched baselines, and the supervised clock provides an explicit estimate of alignment progress.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Nan Lu, Qifan Liang. 2026-10-25. Speaking-Clock Mamba: Phoneme-Conditioned State-Space Modeling for Visual Forced Alignment. https://searcharxiv.com/abs/2611.18667
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.