This article reports on the second GENEA Challenge to benchmark datadriven automatic co-speech gesture generation. Participating teams used the same speech and motion dataset to build gesture-generation systems. Motion generated by all these systems was rendered to video using a standardised visualisation pipeline and evaluated in several large, crowdsourced user studies. Unlike when comparing different research articles, differences in results are here only due to differences betweenmethods, enabling direct comparison between systems. The dataset was based on 18 hours of fullbody motion capture, including fingers, of different persons engaging in a dyadic conversation. Ten teams participated in the challenge across two tiers: full-body and upper-body gesticulation. For each tier, we evaluated both the human-likeness of the gesture motion and its appropriateness for the specific speech signal. Our evaluations decouple human-likeness from gesture appropriateness, which has been a difficult problem in the field. The evaluation results show some synthetic gesture conditions being rated as significantly more human-like than 3D human motion capture. To the best of our knowledge, this has not been demonstrated before. On the other hand, all synthetic motion is found to be vastly less appropriate for the speech than the original motion-capture recordings. We also find that conventional objective metrics do not correlate well with subjective human-likeness ratings in this large evaluation. The one exception is the Frechet gesture distance (FGD), which achieves a Kendall's tau rank correlation of around -0.5. Based on the challenge results we formulate numerous recommendations for system building and evaluation.

It has been found that Transformer-based language models have the ability to perform basic quantitative reasoning. In this paper, we propose a method for studying how these models internally represent numerical data, and use our proposal to analyze the ALBERT family of language models. Specifically, we extract the learned embeddings these models use to represent tokens that correspond to numbers and ordinals, and subject these embeddings to Principal Component Analysis (PCA). PCA results reveal that ALBERT models of different sizes, trained and initialized separately, consistently learn to use the axes of greatest variation to represent the approximate ordering of various numerical concepts. Numerals and their textual counterparts are represented in separate clusters, but increase along the same direction in 2D space. Our findings illustrate that language models, trained purely to model text, can intuit basic mathematical concepts, opening avenues for NLP applications that intersect with quantitative reasoning.

This paper compares three methods for evaluating computer-generated motion behaviour for animated characters: two commonly used direct rating methods and a newly designed questionnaire. The questionnaire is specifically designed to measure the human-likeness, appropriateness, and intelligibility of the generated motion. Furthermore, this study investigates the suitability of these evaluation tools for assessing subtle forms of human behaviour, such as the subdued motion cues shown when listening to someone. This paper reports six user studies, namely studies that directly rate the appropriateness and human-likeness of a computer character's motion, along with studies that instead rely on a questionnaire to measure the quality of the motion. As test data, we used the motion generated by two generative models and recorded human gestures, which served as a gold standard. Our findings indicate that when evaluating gesturing motion, the direct rating of human-likeness and appropriateness is to be preferred over a questionnaire. However, when assessing the subtle motion of a computer character, even the direct rating method yields less conclusive results. Despite demonstrating high internal consistency, our questionnaire proves to be less sensitive than directly rating the quality of the motion. The results provide insights into the evaluation of human motion behaviour and highlight the complexities involved in capturing subtle nuances in nonverbal communication. These findings have implications for the development and improvement of motion generation models and can guide researchers in selecting appropriate evaluation methodologies for specific aspects of human behaviour.

Non-verbal behavior offers significant benefits for embodied agents in human interactions. Despite extensive research on the creation of non-verbal behaviors, the field lacks a standardized benchmarking practice. Researchers rarely compare their findings with previous studies, and when they do, the comparisons are often not aligned with other methodologies. The GENEA Workshop 2024 aims to unite the community to discuss major challenges and solutions, and to determine the most effective ways to advance the field.

We introduce Matcha-TTS, a new encoder-decoder architecture for speedy TTS acoustic modelling, trained using optimal-transport conditional flow matching (OT-CFM). This yields an ODE-based decoder capable of high output quality in fewer synthesis steps than models trained using score matching. Careful design choices additionally ensure each synthesis step is fast to run. The method is probabilistic, non-autoregressive, and learns to speak from scratch without external alignments. Compared to strong pre-trained baseline models, the Matcha-TTS system has the smallest memory footprint, rivals the speed of the fastest model on long utterances, and attains the highest mean opinion score in a listening test.

Converting input symbols to output audio in TTS requires modelling the durations of speech sounds. Leading non-autoregressive (NAR) TTS models treat duration modelling as a regression problem. The same utterance is then spoken with identical timings every time, unlike when a human speaks. Probabilistic models of duration have been proposed, but there is mixed evidence of their benefits. However, prior studies generally only consider speech read aloud, and ignore spontaneous speech, despite the latter being both a more common and a more variable mode of speaking. We compare the effect of conventional deterministic duration modelling to durations sampled from a powerful probabilistic model based on conditional flow matching (OT-CFM), in three different NAR TTS approaches: regression-based, deep generative, and end-to-end. Across four different corpora, stochastic duration modelling improves probabilistic NAR TTS approaches, especially for spontaneous speech.

As text-to-speech technologies achieve remarkable naturalness in read-aloud tasks, there is growing interest in multimodal synthesis of verbal and non-verbal communicative behaviour, such as spontaneous speech and associated body gestures. This paper presents a novel, unified architecture for jointly synthesising speech acoustics and skeleton-based 3D gesture motion from text, trained using optimal-transport conditional flow matching (OT-CFM). The proposed architecture is simpler than the previous state of the art, has a smaller memory footprint, and can capture the joint distribution of speech and gestures, generating both modalities together in one single process. The new training regime, meanwhile, enables better synthesis quality in much fewer steps (network evaluations) than before. Uni- and multimodal subjective tests demonstrate improved speech naturalness, gesture human-likeness, and cross-modal appropriateness compared to existing benchmarks.

Recent work has explored using self-supervised learning (SSL) speech representations such as wav2vec2.0 as the representation medium in standard two-stage TTS, in place of conventionally used mel-spectrograms. It is however unclear which speech SSL is the better fit for TTS, and whether or not the performance differs between read and spontaneous TTS, the later of which is arguably more challenging. This study aims at addressing these questions by testing several speech SSLs, including different layers of the same SSL, in two-stage TTS on both read and spontaneous corpora, while maintaining constant TTS model architecture and training settings. Results from listening tests show that the 9th layer of 12-layer wav2vec2.0 (ASR finetuned) outperforms other tested SSLs and mel-spectrogram, in both read and spontaneous TTS. Our work sheds light on both how speech SSL can readily improve current TTS systems, and how SSLs compare in the challenging generative task of TTS. Audio examples can be found at https://www.speech.kth.se/tts-demos/ssr_tts

Recent work has explored using self-supervised learning(SSL) speech representations such as wav2vec2.0 as the rep-resentation medium in standard two-stage TTS, in place ofconventionally used mel-spectrograms. It is however unclearwhich speech SSL is the better fit for TTS, and whether ornot the performance differs between read and spontaneousTTS, the later of which is arguably more challenging. Thisstudy aims at addressing these questions by testing severalspeech SSLs, including different layers of the same SSL, intwo-stage TTS on both read and spontaneous corpora, whilemaintaining constant TTS model architecture and trainingsettings. Results from listening tests show that the 9th layerof 12-layer wav2vec2.0 (ASR finetuned) outperforms othertested SSLs and mel-spectrogram, in both read and sponta-neous TTS. Our work sheds light on both how speech SSL canreadily improve current TTS systems, and how SSLs comparein the challenging generative task of TTS. Audio examplescan be found at https://www.speech.kth.se/tts-demos/ssr tts