Refine
Document Type
Language
- English (5)
Is part of the Bibliography
- no (5)
Keywords
- German (1)
- audio-visual synchrony (1)
- corpus (1)
- cross-modal learning (1)
- dubbing (1)
- forced alignment (1)
- multi-modal (1)
- single-speaker (1)
- speaker diarization (1)
- speech-lip synchrony (1)
Institute
Begutachtungsstatus
- peer-reviewed (4)
We present a comprehensive analysis of the neural audio-visual synchrony evaluation tool SyncNet. We assess the agreement of SyncNet scores vis-a-vis human perception and whether we can use these as a reliable metric for evaluating audio-visual lip-synchrony in generation tasks with no ground truth reference audio-video pair. We further look into the underlying elements in audio and video which vitally affect synchrony using interpretable explanations from SyncNet predictions and analyse its susceptibility by introducing adversarial noise. SyncNet has been used in numerous papers on visually-grounded text-to-speech for scenarios such as dubbing. We focus on this scenario which features many local asynchronies (something that SyncNet isn’t made for).
We introduce the Merkel Podcast Corpus, an audio-visual-text corpus in German collected from 16 years of (almost) weekly Internet podcasts of former German chancellor Angela Merkel. To the best of our knowledge, this is the first single speaker corpus in the German language consisting of audio, visual and text modalities of comparable size and temporal extent. We describe the methods used with which we have collected and edited the data which involves downloading the videos, transcripts and other metadata, forced alignment, performing active speaker recognition and face detection to finally curate the single speaker dataset consisting of utterances spoken by Angela Merkel. The proposed pipeline is general and can be used to curate other datasets of similar nature, such as talk show contents. Through various statistical analyses and applications of the dataset in talking face generation and TTS, we show the utility of the dataset. We argue that it is a valuable contribution to the research community, in particular, due to its realistic and challenging material at the boundary between prepared and spontaneous speech.
Accepted at LREC 2022
Dubbing is the art of finding a translation from a source into a target language that can be lip-synchronously revoiced, i. e., that makes the target language speech appear as if it was spoken by the very actors all along. Lip synchrony is essential for the full-fledged reception of foreign audiovisual media, such as movies and series, as violated constraints of synchrony between video (lips) and audio (speech) lead to cognitive dissonance and reduce the perceptual quality. Of course, synchrony constraints only apply to the translation when the speaker's lips are visible on screen. Therefore, deciding whether to apply synchrony constraints requires an automatic method for detecting whether an actor's lips are visible on screen for a given stretch of speech or not. In this paper, we attempt, for the first time, to classify on- from off-screen speech based on a corpus of real-world television material that has been annotated word-by-word for the visibility of talking lips on screen. We present classification experiments in which we classify
Dubbing has two shades; synchronisation constraints are applied only when the actor’s mouth is visible on screen, while the translation is unconstrained for off-screen dubbing. Consequently, different synchronisation requirements, and therefore translation strategies, are applied depending on the type of dubbing. In this work, we manually annotate an existing dubbing corpus (Heroes) for this dichotomy. We show that, even though we did not observe distinctive features between on- and off-screen dubbing at the textual level, on-screen dubbing is more difficult for MT (-4 BLEU points). Moreover, synchronisation constraints dramatically decrease translation quality for off-screen dubbing. We conclude that, distinguishing between on-screen and off-screen dubbing is necessary for determining successful strategies for dubbing-customised Machine Translation.
In this work we assess whether there is information in pauses in-between utterances of the same or different speakers that are predictive of the following speaker’s utterance. We present models that connect a person’s visual features before they speak to their upcoming utterance. In our experiments we find that outof-the-box pre-trained models can already reach a better-than-chance performance in correlating video embeddings to utterance embeddings. In contrast, models that attempt to predict the first word after the pause do not outperform a unigram model, indicating that our models do not read lips (based e.g. on co-articulation effects) but rather capture more fundamental aspects of the upcoming utterance.