Robust temporal alignment of spontaneous and dubbed speech and its application for automatic dialogue replacement

In this paper, we present a robust system for the temporal alignment of 2 renditions of the same speech utterance. The system operates in 2 steps: during analysis, the timing relationships between the speech segments of the utterance that serves as a timing reference and the corresponding speech segments in the replacement utterance are measured by means of a dedicated dynamic time warping algorithm. The obtained warping paths are then processed and used to synthesize a high-quality speech utterance that is time-aligned with the reference. Subjective audio-visual listening tests performed within the context of a difficult Automatic Dialogue Replacement task demonstrated that the proposed system achieves a significant improvement compared to the industry-standard benchmark, both in terms of achieved lip-synchronization accuracy as well as in overall sound quality of the synthesized utterances.