Gustav Eje Henter

Rajmund Nagy, Hendric Voss, Thanh Hoang-Minh et al.

We review human evaluation practices in automated, speech-driven 3D gesture generation and find a lack of standardisation and frequent use of flawed experimental setups. This leads to a situation where it is impossible to know how different methods compare, or what the state of the art is. In order to address common shortcomings of evaluation design, and to standardise future user studies in gesture-generation works, we introduce a detailed human evaluation protocol for the widely-used BEAT2 motion-capture dataset. Using this protocol, we conduct large-scale crowdsourced evaluation to rank six recent gesture-generation models -- each trained by its original authors -- across two key evaluation dimensions: motion realism and speech-gesture alignment. Our results provide strong evidence that 1) newer models do not consistently outperform earlier approaches; 2) published claims of high motion realism or speech-gesture alignment may not hold up under rigorous evaluation; and 3) the field must adopt disentangled assessments of motion quality and multimodal alignment for accurate benchmarking in order to make progress. Finally, in order to drive standardisation and enable new evaluation research, we will release five hours of synthetic motion from the benchmarked models; over 750 rendered video stimuli from the user studies -- enabling new evaluations without model reimplementation required -- alongside our open-source rendering script, and the 16,000 pairwise human preference votes collected for our benchmark.

Simon Alexanderson, Rajmund Nagy, Jonas Beskow et al.

Diffusion models have experienced a surge of interest as highly expressive yet efficiently trainable probabilistic models. We show that these models are an excellent fit for synthesising human motion that co-occurs with audio, e.g., dancing and co-speech gesticulation, since motion is complex and highly ambiguous given audio, calling for a probabilistic description. Specifically, we adapt the DiffWave architecture to model 3D pose sequences, putting Conformers in place of dilated convolutions for improved modelling power. We also demonstrate control over motion style, using classifier-free guidance to adjust the strength of the stylistic expression. Experiments on gesture and dance generation confirm that the proposed method achieves top-of-the-line motion quality, with distinctive styles whose expression can be made more or less pronounced. We also synthesise path-driven locomotion using the same model architecture. Finally, we generalise the guidance procedure to obtain product-of-expert ensembles of diffusion models and demonstrate how these may be used for, e.g., style interpolation, a contribution we believe is of independent interest. See https://www.speech.kth.se/research/listen-denoise-action/ for video examples, data, and code.

Shivam Mehta, Ruibo Tu, Jonas Beskow et al.

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 models on long utterances, and attains the highest mean opinion score in a listening test. Please see https://shivammehta25.github.io/Matcha-TTS/ for audio examples, code, and pre-trained models.

Simbarashe Nyatsanga, Taras Kucherenko, Chaitanya Ahuja et al.

Gestures that accompany speech are an essential part of natural and efficient embodied human communication. The automatic generation of such co-speech gestures is a long-standing problem in computer animation and is considered an enabling technology in film, games, virtual social spaces, and for interaction with social robots. The problem is made challenging by the idiosyncratic and non-periodic nature of human co-speech gesture motion, and by the great diversity of communicative functions that gestures encompass. Gesture generation has seen surging interest recently, owing to the emergence of more and larger datasets of human gesture motion, combined with strides in deep-learning-based generative models, that benefit from the growing availability of data. This review article summarizes co-speech gesture generation research, with a particular focus on deep generative models. First, we articulate the theory describing human gesticulation and how it complements speech. Next, we briefly discuss rule-based and classical statistical gesture synthesis, before delving into deep learning approaches. We employ the choice of input modalities as an organizing principle, examining systems that generate gestures from audio, text, and non-linguistic input. We also chronicle the evolution of the related training data sets in terms of size, diversity, motion quality, and collection method. Finally, we identify key research challenges in gesture generation, including data availability and quality; producing human-like motion; grounding the gesture in the co-occurring speech in interaction with other speakers, and in the environment; performing gesture evaluation; and integration of gesture synthesis into applications. We highlight recent approaches to tackling the various key challenges, as well as the limitations of these approaches, and point toward areas of future development.

Youngwoo Yoon, Pieter Wolfert, Taras Kucherenko et al.

This paper reports on the second GENEA Challenge to benchmark data-driven 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 papers, differences in results are here only due to differences between methods, enabling direct comparison between systems. This year's dataset was based on 18 hours of full-body motion capture, including fingers, of different persons engaging in 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 previously was a major challenge in the field. The evaluation results are a revolution, and a revelation. Some synthetic conditions are rated as significantly more human-like than human motion capture. To the best of our knowledge, this has never been shown before on a high-fidelity avatar. On the other hand, all synthetic motion is found to be vastly less appropriate for the speech than the original motion-capture recordings. Additional material is available via the project website at https://youngwoo-yoon.github.io/GENEAchallenge2022/

Taras Kucherenko, Rajmund Nagy, Youngwoo Yoon et al.

This paper reports on the GENEA Challenge 2023, in which participating teams built speech-driven gesture-generation systems using the same speech and motion dataset, followed by a joint evaluation. This year's challenge provided data on both sides of a dyadic interaction, allowing teams to generate full-body motion for an agent given its speech (text and audio) and the speech and motion of the interlocutor. We evaluated 12 submissions and 2 baselines together with held-out motion-capture data in several large-scale user studies. The studies focused on three aspects: 1) the human-likeness of the motion, 2) the appropriateness of the motion for the agent's own speech whilst controlling for the human-likeness of the motion, and 3) the appropriateness of the motion for the behaviour of the interlocutor in the interaction, using a setup that controls for both the human-likeness of the motion and the agent's own speech. We found a large span in human-likeness between challenge submissions, with a few systems rated close to human mocap. Appropriateness seems far from being solved, with most submissions performing in a narrow range slightly above chance, far behind natural motion. The effect of the interlocutor is even more subtle, with submitted systems at best performing barely above chance. Interestingly, a dyadic system being highly appropriate for agent speech does not necessarily imply high appropriateness for the interlocutor. Additional material is available via the project website at https://svito-zar.github.io/GENEAchallenge2023/ .

Taras Kucherenko, Pieter Wolfert, Youngwoo Yoon et al.

This paper reports on the second GENEA Challenge to benchmark data-driven 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 papers, differences in results are here only due to differences between methods, enabling direct comparison between systems. The dataset was based on 18 hours of full-body 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 Fréchet 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.

Harm Lameris, Shivam Mehta, Gustav Eje Henter et al.

Spontaneous speech has many affective and pragmatic functions that are interesting and challenging to model in TTS. However, the presence of reduced articulation, fillers, repetitions, and other disfluencies in spontaneous speech make the text and acoustics less aligned than in read speech, which is problematic for attention-based TTS. We propose a TTS architecture that can rapidly learn to speak from small and irregular datasets, while also reproducing the diversity of expressive phenomena present in spontaneous speech. Specifically, we add utterance-level prosody control to an existing neural HMM-based TTS system which is capable of stable, monotonic alignments for spontaneous speech. We objectively evaluate control accuracy and perform perceptual tests that demonstrate that prosody control does not degrade synthesis quality. To exemplify the power of combining prosody control and ecologically valid data for reproducing intricate spontaneous speech phenomena, we evaluate the system's capability of synthesizing two types of creaky voice. Audio samples are available at https://www.speech.kth.se/tts-demos/prosodic-hmm/

Shivam Mehta, Ambika Kirkland, Harm Lameris et al.

Neural HMMs are a type of neural transducer recently proposed for sequence-to-sequence modelling in text-to-speech. They combine the best features of classic statistical speech synthesis and modern neural TTS, requiring less data and fewer training updates, and are less prone to gibberish output caused by neural attention failures. In this paper, we combine neural HMM TTS with normalising flows for describing the highly non-Gaussian distribution of speech acoustics. The result is a powerful, fully probabilistic model of durations and acoustics that can be trained using exact maximum likelihood. Experiments show that a system based on our proposal needs fewer updates than comparable methods to produce accurate pronunciations and a subjective speech quality close to natural speech. Please see https://shivammehta25.github.io/OverFlow/ for audio examples and code.

Shivam Mehta, Siyang Wang, Simon Alexanderson et al.

With read-aloud speech synthesis achieving high naturalness scores, there is a growing research interest in synthesising spontaneous speech. However, human spontaneous face-to-face conversation has both spoken and non-verbal aspects (here, co-speech gestures). Only recently has research begun to explore the benefits of jointly synthesising these two modalities in a single system. The previous state of the art used non-probabilistic methods, which fail to capture the variability of human speech and motion, and risk producing oversmoothing artefacts and sub-optimal synthesis quality. We present the first diffusion-based probabilistic model, called Diff-TTSG, that jointly learns to synthesise speech and gestures together. Our method can be trained on small datasets from scratch. Furthermore, we describe a set of careful uni- and multi-modal subjective tests for evaluating integrated speech and gesture synthesis systems, and use them to validate our proposed approach. Please see https://shivammehta25.github.io/Diff-TTSG/ for video examples, data, and code.

Siyang Wang, Gustav Eje Henter, Joakim Gustafson et al.

Self-supervised learning (SSL) speech representations learned from large amounts of diverse, mixed-quality speech data without transcriptions are gaining ground in many speech technology applications. Prior work has shown that SSL is an effective intermediate representation in two-stage text-to-speech (TTS) for both read and spontaneous speech. However, it is still not clear which SSL and which layer from each SSL model is most suited for spontaneous TTS. We address this shortcoming by extending the scope of comparison for SSL in spontaneous TTS to 6 different SSLs and 3 layers within each SSL. Furthermore, SSL has also shown potential in predicting the mean opinion scores (MOS) of synthesized speech, but this has only been done in read-speech MOS prediction. We extend an SSL-based MOS prediction framework previously developed for scoring read speech synthesis and evaluate its performance on synthesized spontaneous speech. All experiments are conducted twice on two different spontaneous corpora in order to find generalizable trends. Overall, we present comprehensive experimental results on the use of SSL in spontaneous TTS and MOS prediction to further quantify and understand how SSL can be used in spontaneous TTS. Audios samples: https://www.speech.kth.se/tts-demos/sp_ssl_tts

Carlos Puerto-Santana, Concha Bielza, Pedro Larrañaga et al.

Traditional hidden Markov models have been a useful tool to understand and model stochastic dynamic data; in the case of non-Gaussian data, models such as mixture of Gaussian hidden Markov models can be used. However, these suffer from the computation of precision matrices and have a lot of unnecessary parameters. As a consequence, such models often perform better when it is assumed that all variables are independent, a hypothesis that may be unrealistic. Hidden Markov models based on kernel density estimation are also capable of modeling non-Gaussian data, but they assume independence between variables. In this article, we introduce a new hidden Markov model based on kernel density estimation, which is capable of capturing kernel dependencies using context-specific Bayesian networks. The proposed model is described, together with a learning algorithm based on the expectation-maximization algorithm. Additionally, the model is compared to related HMMs on synthetic and real data. From the results, the benefits in likelihood and classification accuracy from the proposed model are quantified and analyzed.

Siyang Wang, Gustav Eje Henter, Joakim Gustafson et al.

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

Shivam Mehta, Ruibo Tu, Simon Alexanderson et al.

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. Please see https://shivammehta25.github.io/Match-TTSG/ for video examples and code.

Cassia Valentini-Botinhao, Manuel Sam Ribeiro, Oliver Watts et al.

Automatically predicting the outcome of subjective listening tests is a challenging task. Ratings may vary from person to person even if preferences are consistent across listeners. While previous work has focused on predicting listeners' ratings (mean opinion scores) of individual stimuli, we focus on the simpler task of predicting subjective preference given two speech stimuli for the same text. We propose a model based on anti-symmetric twin neural networks, trained on pairs of waveforms and their corresponding preference scores. We explore both attention and recurrent neural nets to account for the fact that stimuli in a pair are not time aligned. To obtain a large training set we convert listeners' ratings from MUSHRA tests to values that reflect how often one stimulus in the pair was rated higher than the other. Specifically, we evaluate performance on data obtained from twelve MUSHRA evaluations conducted over five years, containing different TTS systems, built from data of different speakers. Our results compare favourably to a state-of-the-art model trained to predict MOS scores.

Nikita Torgashov, Gustav Eje Henter, Gabriel Skantze

Full-stream text-to-speech (TTS) for interactive systems must start speaking with minimal delay while remaining controllable as text arrives incrementally. We present VoXtream2, a zero-shot full-stream TTS model with dynamic speaking-rate control that can be updated mid-utterance on the fly. VoXtream2 combines a distribution matching mechanism over duration states with classifier-free guidance across conditioning signals to improve controllability and synthesis quality. Prompt-text masking enables textless audio prompting, removing the need for prompt transcription. Across standard zero-shot benchmarks and a dedicated speaking-rate test set, VoXtream2 achieves competitive objective and subjective results against public baselines despite a smaller model and less training data. In full-stream mode, it runs 4 times faster than real time with 74 ms first-packet latency on a consumer GPU.

Mohammed Al-Jaff, Giovanni Luca Marchetti, Michael C Welle et al.

Idempotent Generative Networks (IGNs) are deep generative models that also function as local data manifold projectors, mapping arbitrary inputs back onto the manifold. They are trained to act as identity operators on the data and as idempotent operators off the data manifold. However, IGNs suffer from mode collapse, mode dropping, and training instability due to their objectives, which contain adversarial components and can cause the model to cover the data manifold only partially -- an issue shared with generative adversarial networks. We introduce Non-Adversarial Idempotent Generative Networks (NAIGNs) to address these issues. Our loss function combines reconstruction with the non-adversarial generative objective of Implicit Maximum Likelihood Estimation (IMLE). This improves on IGN's ability to restore corrupted data and generate new samples that closely match the data distribution. We moreover demonstrate that NAIGNs implicitly learn the distance field to the data manifold, as well as an energy-based model.

Shree Harsha Bokkahalli Satish, Christoph Minixhofer, Maria Teleki et al.

Speech Large Language Models (SpeechLLMs) process spoken input directly, retaining cues such as accent and perceived gender that were previously removed in cascaded pipelines. This introduces speaker identity dependent variation in responses. We present a large-scale intersectional evaluation of accent and gender bias in three SpeechLLMs using 2,880 controlled interactions across six English accents and two gender presentations, keeping linguistic content constant through voice cloning. Using pointwise LLM-judge ratings, pairwise comparisons, and Best-Worst Scaling with human validation, we detect consistent disparities. Eastern European-accented speech receives lower helpfulness scores, particularly for female-presenting voices. The bias is implicit: responses remain polite but differ in helpfulness. While LLM judges capture the directional trend of these biases, human evaluators exhibit significantly higher sensitivity, uncovering sharper intersectional disparities.

Ruibo Tu, Zineb Senane, Lele Cao et al.

Tabular synthesis models remain ineffective at capturing complex dependencies, and the quality of synthetic data is still insufficient for comprehensive downstream tasks, such as prediction under distribution shifts, automated decision-making, and cross-table understanding. A major challenge is the lack of prior knowledge about underlying structures and high-order relationships in tabular data. We argue that a systematic evaluation on high-order structural information for tabular data synthesis is the first step towards solving the problem. In this paper, we introduce high-order structural causal information as natural prior knowledge and provide a benchmark framework for the evaluation of tabular synthesis models. The framework allows us to generate benchmark datasets with a flexible range of data generation processes and to train tabular synthesis models using these datasets for further evaluation. We propose multiple benchmark tasks, high-order metrics, and causal inference tasks as downstream tasks for evaluating the quality of synthetic data generated by the trained models. Our experiments demonstrate to leverage the benchmark framework for evaluating the model capability of capturing high-order structural causal information. Furthermore, our benchmarking results provide an initial assessment of state-of-the-art tabular synthesis models. They have clearly revealed significant gaps between ideal and actual performance and how baseline methods differ. Our benchmark framework is available at URL https://github.com/TURuibo/CauTabBench.

Ruibo Tu, Hedvig Kjellström, Gustav Eje Henter et al.

Causal reasoning capabilities are essential for large language models (LLMs) in a wide range of applications, such as education and healthcare. But there is still a lack of benchmarks for a better understanding of such capabilities. Current LLM benchmarks are mainly based on conversational tasks, academic math tests, and coding tests. Such benchmarks evaluate LLMs in well-regularized settings, but they are limited in assessing the skills and abilities to solve real-world problems. In this work, we provide a benchmark, named by CARL-GT, which evaluates CAusal Reasoning capabilities of large Language models using Graphs and Tabular data. The benchmark has a diverse range of tasks for evaluating LLMs from causal graph reasoning, knowledge discovery, and decision-making aspects. In addition, effective zero-shot learning prompts are developed for the tasks. In our experiments, we leverage the benchmark for evaluating open-source LLMs and provide a detailed comparison of LLMs for causal reasoning abilities. We found that LLMs are still weak in casual reasoning, especially with tabular data to discover new insights. Furthermore, we investigate and discuss the relationships of different benchmark tasks by analyzing the performance of LLMs. The experimental results show that LLMs have different strength over different tasks and that their performance on tasks in different categories, i.e., causal graph reasoning, knowledge discovery, and decision-making, shows stronger correlation than tasks in the same category.

Patrik Jonell, Taras Kucherenko, Gustav Eje Henter et al.

To enable more natural face-to-face interactions, conversational agents need to adapt their behavior to their interlocutors. One key aspect of this is generation of appropriate non-verbal behavior for the agent, for example facial gestures, here defined as facial expressions and head movements. Most existing gesture-generating systems do not utilize multi-modal cues from the interlocutor when synthesizing non-verbal behavior. Those that do, typically use deterministic methods that risk producing repetitive and non-vivid motions. In this paper, we introduce a probabilistic method to synthesize interlocutor-aware facial gestures - represented by highly expressive FLAME parameters - in dyadic conversations. Our contributions are: a) a method for feature extraction from multi-party video and speech recordings, resulting in a representation that allows for independent control and manipulation of expression and speech articulation in a 3D avatar; b) an extension to MoGlow, a recent motion-synthesis method based on normalizing flows, to also take multi-modal signals from the interlocutor as input and subsequently output interlocutor-aware facial gestures; and c) a subjective evaluation assessing the use and relative importance of the input modalities. The results show that the model successfully leverages the input from the interlocutor to generate more appropriate behavior. Videos, data, and code available at: https://jonepatr.github.io/lets_face_it.

Shivam Mehta, Anna Deichler, Jim O'Regan et al.

Although humans engaged in face-to-face conversation simultaneously communicate both verbally and non-verbally, methods for joint and unified synthesis of speech audio and co-speech 3D gesture motion from text are a new and emerging field. These technologies hold great promise for more human-like, efficient, expressive, and robust synthetic communication, but are currently held back by the lack of suitably large datasets, as existing methods are trained on parallel data from all constituent modalities. Inspired by student-teacher methods, we propose a straightforward solution to the data shortage, by simply synthesising additional training material. Specifically, we use unimodal synthesis models trained on large datasets to create multimodal (but synthetic) parallel training data, and then pre-train a joint synthesis model on that material. In addition, we propose a new synthesis architecture that adds better and more controllable prosody modelling to the state-of-the-art method in the field. Our results confirm that pre-training on large amounts of synthetic data improves the quality of both the speech and the motion synthesised by the multimodal model, with the proposed architecture yielding further benefits when pre-trained on the synthetic data. See https://shivammehta25.github.io/MAGI/ for example output.

Ulme Wennberg, Gustav Eje Henter

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.

Shree Harsha Bokkahalli Satish, Gustav Eje Henter, Éva Székely

Recent work in benchmarking bias and fairness in speech large language models (SpeechLLMs) has relied heavily on multiple-choice question answering (MCQA) formats. The model is tasked to choose between stereotypical, anti-stereotypical, or neutral/irrelevant answers given an input speech prompt and an optional text prompt. Such MCQA benchmarks implicitly assume that model performance is consistent across other MCQA tasks, voices, and other task formats such as more realistic, long-form evaluations. In this paper, we probe that assumption. We fine-tune three SpeechLLMs using LoRA adapters to induce specific MCQA behaviours: preference for stereotypical, anti-stereotypical, or neutral/uncertain answers. We then evaluate whether these behaviours generalise to another, distinct MCQA benchmark, and more critically to long-form, creative generation tasks. Our results show that performance on MCQA bias benchmarks fails to reliably predict performances across other MCQA benchmarks, and more importantly across long-form tasks. We conclude that current MCQA bias benchmarks show limited evidence of cross-task generalisation in the speech domain, and also propose an evaluation suite for measuring behaviour transferability in future models and benchmarks.

Shree Harsha Bokkahalli Satish, Harm Lameris, Olivier Perrotin et al.

Speech Continuation (SC) is the task of generating a coherent extension of a spoken prompt while preserving both semantic context and speaker identity. Because SC is constrained to a single audio stream, it offers a more direct setting for probing biases in speech foundation models than dialogue does. In this work we present the first systematic evaluation of bias in SC, investigating how gender and phonation type (breathy, creaky, end-creak) affect continuation behaviour. We evaluate three recent models: SpiritLM (base and expressive), VAE-GSLM, and SpeechGPT across speaker similarity, voice quality preservation, and text-based bias metrics. Results show that while both speaker similarity and coherence remain a challenge, textual evaluations reveal significant model and gender interactions: once coherence is sufficiently high (for VAE-GSLM), gender effects emerge on text-metrics such as agency and sentence polarity. In addition, continuations revert toward modal phonation more strongly for female prompts than for male ones, revealing a systematic voice-quality bias. These findings highlight SC as a controlled probe of socially relevant representational biases in speech foundation models, and suggest that it will become an increasingly informative diagnostic as continuation quality improves.

Nikita Torgashov, Gustav Eje Henter, Gabriel Skantze

We present VoXtream, a fully autoregressive, zero-shot streaming text-to-speech (TTS) system for real-time use that begins speaking from the first word. VoXtream directly maps incoming phonemes to audio tokens using a monotonic alignment scheme and a dynamic look-ahead that does not delay onset. Built around an incremental phoneme transformer, a temporal transformer predicting semantic and duration tokens, and a depth transformer producing acoustic tokens, VoXtream achieves, to our knowledge, the lowest initial delay among publicly available streaming TTS: 102 ms on GPU. Despite being trained on a mid-scale 9k-hour corpus, it matches or surpasses larger baselines on several metrics, while delivering competitive quality in both output- and full-streaming settings. Demo and code are available at https://herimor.github.io/voxtream.

Gustavo Teodoro Döhler Beck, Ulme Wennberg, Zofia Malisz et al.

Deep learning has revolutionised synthetic speech quality. However, it has thus far delivered little value to the speech science community. The new methods do not meet the controllability demands that practitioners in this area require e.g.: in listening tests with manipulated speech stimuli. Instead, control of different speech properties in such stimuli is achieved by using legacy signal-processing methods. This limits the range, accuracy, and speech quality of the manipulations. Also, audible artefacts have a negative impact on the methodological validity of results in speech perception studies. This work introduces a system capable of manipulating speech properties through learning rather than design. The architecture learns to control arbitrary speech properties and leverages progress in neural vocoders to obtain realistic output. Experiments with copy synthesis and manipulation of a small set of core speech features (pitch, formants, and voice quality measures) illustrate the promise of the approach for producing speech stimuli that have accurate control and high perceptual quality.

Shivam Mehta, Éva Székely, Jonas Beskow et al.

Neural sequence-to-sequence TTS has achieved significantly better output quality than statistical speech synthesis using HMMs. However, neural TTS is generally not probabilistic and uses non-monotonic attention. Attention failures increase training time and can make synthesis babble incoherently. This paper describes how the old and new paradigms can be combined to obtain the advantages of both worlds, by replacing attention in neural TTS with an autoregressive left-right no-skip hidden Markov model defined by a neural network. Based on this proposal, we modify Tacotron 2 to obtain an HMM-based neural TTS model with monotonic alignment, trained to maximise the full sequence likelihood without approximation. We also describe how to combine ideas from classical and contemporary TTS for best results. The resulting example system is smaller and simpler than Tacotron 2, and learns to speak with fewer iterations and less data, whilst achieving comparable naturalness prior to the post-net. Our approach also allows easy control over speaking rate.

Siyang Wang, Simon Alexanderson, Joakim Gustafson et al.

Text-to-speech and co-speech gesture synthesis have until now been treated as separate areas by two different research communities, and applications merely stack the two technologies using a simple system-level pipeline. This can lead to modeling inefficiencies and may introduce inconsistencies that limit the achievable naturalness. We propose to instead synthesize the two modalities in a single model, a new problem we call integrated speech and gesture synthesis (ISG). We also propose a set of models modified from state-of-the-art neural speech-synthesis engines to achieve this goal. We evaluate the models in three carefully-designed user studies, two of which evaluate the synthesized speech and gesture in isolation, plus a combined study that evaluates the models like they will be used in real-world applications -- speech and gesture presented together. The results show that participants rate one of the proposed integrated synthesis models as being as good as the state-of-the-art pipeline system we compare against, in all three tests. The model is able to achieve this with faster synthesis time and greatly reduced parameter count compared to the pipeline system, illustrating some of the potential benefits of treating speech and gesture synthesis together as a single, unified problem. Videos and code are available on our project page at https://swatsw.github.io/isg_icmi21/

Taras Kucherenko, Rajmund Nagy, Michael Neff et al.

Embodied conversational agents benefit from being able to accompany their speech with gestures. Although many data-driven approaches to gesture generation have been proposed in recent years, it is still unclear whether such systems can consistently generate gestures that convey meaning. We investigate which gesture properties (phase, category, and semantics) can be predicted from speech text and/or audio using contemporary deep learning. In extensive experiments, we show that gesture properties related to gesture meaning (semantics and category) are predictable from text features (time-aligned FastText embeddings) alone, but not from prosodic audio features, while rhythm-related gesture properties (phase) on the other hand can be predicted from audio features better than from text. These results are encouraging as they indicate that it is possible to equip an embodied agent with content-wise meaningful co-speech gestures using a machine-learning model.

Anubhab Ghosh, Antoine Honoré, Dong Liu et al.

In pursuit of explainability, we develop generative models for sequential data. The proposed models provide state-of-the-art classification results and robust performance for speech phone classification. We combine modern neural networks (normalizing flows) and traditional generative models (hidden Markov models - HMMs). Normalizing flow-based mixture models (NMMs) are used to model the conditional probability distribution given the hidden state in the HMMs. Model parameters are learned through judicious combinations of time-tested Bayesian learning methods and contemporary neural network learning methods. We mainly combine expectation-maximization (EM) and mini-batch gradient descent. The proposed generative models can compute likelihood of a data and hence directly suitable for maximum-likelihood (ML) classification approach. Due to structural flexibility of HMMs, we can use different normalizing flow models. This leads to different types of HMMs providing diversity in data modeling capacity. The diversity provides an opportunity for easy decision fusion from different models. For a standard speech phone classification setup involving 39 phones (classes) and the TIMIT dataset, we show that the use of standard features called mel-frequency-cepstral-coeffcients (MFCCs), the proposed generative models, and the decision fusion together can achieve $86.6\%$ accuracy by generative training only. This result is close to state-of-the-art results, for examples, $86.2\%$ accuracy of PyTorch-Kaldi toolkit [1], and $85.1\%$ accuracy using light gated recurrent units [2]. We do not use any discriminative learning approach and related sophisticated features in this article.

Taras Kucherenko, Rajmund Nagy, Patrik Jonell et al.

We propose a new framework for gesture generation, aiming to allow data-driven approaches to produce more semantically rich gestures. Our approach first predicts whether to gesture, followed by a prediction of the gesture properties. Those properties are then used as conditioning for a modern probabilistic gesture-generation model capable of high-quality output. This empowers the approach to generate gestures that are both diverse and representational. Follow-ups and more information can be found on the project page: https://svito-zar.github.io/speech2properties2gestures/ .

Guillermo Valle-Pérez, Gustav Eje Henter, Jonas Beskow et al.

Dance requires skillful composition of complex movements that follow rhythmic, tonal and timbral features of music. Formally, generating dance conditioned on a piece of music can be expressed as a problem of modelling a high-dimensional continuous motion signal, conditioned on an audio signal. In this work we make two contributions to tackle this problem. First, we present a novel probabilistic autoregressive architecture that models the distribution over future poses with a normalizing flow conditioned on previous poses as well as music context, using a multimodal transformer encoder. Second, we introduce the currently largest 3D dance-motion dataset, obtained with a variety of motion-capture technologies, and including both professional and casual dancers. Using this dataset, we compare our new model against two baselines, via objective metrics and a user study, and show that both the ability to model a probability distribution, as well as being able to attend over a large motion and music context are necessary to produce interesting, diverse, and realistic dance that matches the music.

Ulme Wennberg, Gustav Eje Henter

Mechanisms for encoding positional information are central for transformer-based language models. In this paper, we analyze the position embeddings of existing language models, finding strong evidence of translation invariance, both for the embeddings themselves and for their effect on self-attention. The degree of translation invariance increases during training and correlates positively with model performance. Our findings lead us to propose translation-invariant self-attention (TISA), which accounts for the relative position between tokens in an interpretable fashion without needing conventional position embeddings. Our proposal has several theoretical advantages over existing position-representation approaches. Experiments show that it improves on regular ALBERT on GLUE tasks, while only adding orders of magnitude less positional parameters.

Taras Kucherenko, Patrik Jonell, Youngwoo Yoon et al.

Co-speech gestures, gestures that accompany speech, play an important role in human communication. Automatic co-speech gesture generation is thus a key enabling technology for embodied conversational agents (ECAs), since humans expect ECAs to be capable of multi-modal communication. Research into gesture generation is rapidly gravitating towards data-driven methods. Unfortunately, individual research efforts in the field are difficult to compare: there are no established benchmarks, and each study tends to use its own dataset, motion visualisation, and evaluation methodology. To address this situation, we launched the GENEA Challenge, a gesture-generation challenge wherein participating teams built automatic gesture-generation systems on a common dataset, and the resulting systems were evaluated in parallel in a large, crowdsourced user study using the same motion-rendering pipeline. Since differences in evaluation outcomes between systems now are solely attributable to differences between the motion-generation methods, this enables benchmarking recent approaches against one another in order to get a better impression of the state of the art in the field. This paper reports on the purpose, design, results, and implications of our challenge.

Anubhab Ghosh, Antoine Honoré, Dong Liu et al.

We test the robustness of a maximum-likelihood (ML) based classifier where sequential data as observation is corrupted by noise. The hypothesis is that a generative model, that combines the state transitions of a hidden Markov model (HMM) and the neural network based probability distributions for the hidden states of the HMM, can provide a robust classification performance. The combined model is called normalizing-flow mixture model based HMM (NMM-HMM). It can be trained using a combination of expectation-maximization (EM) and backpropagation. We verify the improved robustness of NMM-HMM classifiers in an application to speech recognition.

Patrik Jonell, Youngwoo Yoon, Pieter Wolfert et al.

In many research areas, for example motion and gesture generation, objective measures alone do not provide an accurate impression of key stimulus traits such as perceived quality or appropriateness. The gold standard is instead to evaluate these aspects through user studies, especially subjective evaluations of video stimuli. Common evaluation paradigms either present individual stimuli to be scored on Likert-type scales, or ask users to compare and rate videos in a pairwise fashion. However, the time and resources required for such evaluations scale poorly as the number of conditions to be compared increases. Building on standards used for evaluating the quality of multimedia codecs, this paper instead introduces a framework for granular rating of multiple comparable videos in parallel. This methodology essentially analyses all condition pairs at once. Our contributions are 1) a proposed framework, called HEMVIP, for parallel and granular evaluation of multiple video stimuli and 2) a validation study confirming that results obtained using the tool are in close agreement with results of prior studies using conventional multiple pairwise comparisons.

Simon Alexanderson, Éva Székely, Gustav Eje Henter et al.

Embodied human communication encompasses both verbal (speech) and non-verbal information (e.g., gesture and head movements). Recent advances in machine learning have substantially improved the technologies for generating synthetic versions of both of these types of data: On the speech side, text-to-speech systems are now able to generate highly convincing, spontaneous-sounding speech using unscripted speech audio as the source material. On the motion side, probabilistic motion-generation methods can now synthesise vivid and lifelike speech-driven 3D gesticulation. In this paper, we put these two state-of-the-art technologies together in a coherent fashion for the first time. Concretely, we demonstrate a proof-of-concept system trained on a single-speaker audio and motion-capture dataset, that is able to generate both speech and full-body gestures together from text input. In contrast to previous approaches for joint speech-and-gesture generation, we generate full-body gestures from speech synthesis trained on recordings of spontaneous speech from the same person as the motion-capture data. We illustrate our results by visualising gesture spaces and text-speech-gesture alignments, and through a demonstration video at https://simonalexanderson.github.io/IVA2020 .

Moein Sorkhei, Gustav Eje Henter, Hedvig Kjellström

Autonomous agents, such as driverless cars, require large amounts of labeled visual data for their training. A viable approach for acquiring such data is training a generative model with collected real data, and then augmenting the collected real dataset with synthetic images from the model, generated with control of the scene layout and ground truth labeling. In this paper we propose Full-Glow, a fully conditional Glow-based architecture for generating plausible and realistic images of novel street scenes given a semantic segmentation map indicating the scene layout. Benchmark comparisons show our model to outperform recent works in terms of the semantic segmentation performance of a pretrained PSPNet. This indicates that images from our model are, to a higher degree than from other models, similar to real images of the same kinds of scenes and objects, making them suitable as training data for a visual semantic segmentation or object recognition system.

Taras Kucherenko, Dai Hasegawa, Naoshi Kaneko et al.

This paper presents a novel framework for speech-driven gesture production, applicable to virtual agents to enhance human-computer interaction. Specifically, we extend recent deep-learning-based, data-driven methods for speech-driven gesture generation by incorporating representation learning. Our model takes speech as input and produces gestures as output, in the form of a sequence of 3D coordinates. We provide an analysis of different representations for the input (speech) and the output (motion) of the network by both objective and subjective evaluations. We also analyse the importance of smoothing of the produced motion. Our results indicated that the proposed method improved on our baseline in terms of objective measures. For example, it better captured the motion dynamics and better matched the motion-speed distribution. Moreover, we performed user studies on two different datasets. The studies confirmed that our proposed method is perceived as more natural than the baseline, although the difference in the studies was eliminated by appropriate post-processing: hip-centering and smoothing. We conclude that it is important to take both motion representation and post-processing into account when designing an automatic gesture-production method.

Simon Alexanderson, Gustav Eje Henter

Normalising flows are tractable probabilistic models that leverage the power of deep learning to describe a wide parametric family of distributions, all while remaining trainable using maximum likelihood. We discuss how these methods can be further improved based on insights from robust (in particular, resistant) statistics. Specifically, we propose to endow flow-based models with fat-tailed latent distributions such as multivariate Student's $t$, as a simple drop-in replacement for the Gaussian distribution used by conventional normalising flows. While robustness brings many advantages, this paper explores two of them: 1) We describe how using fatter-tailed base distributions can give benefits similar to gradient clipping, but without compromising the asymptotic consistency of the method. 2) We also discuss how robust ideas lead to models with reduced generalisation gap and improved held-out data likelihood. Experiments on several different datasets confirm the efficacy of the proposed approach in both regards.

Taras Kucherenko, Patrik Jonell, Sanne van Waveren et al.

During speech, people spontaneously gesticulate, which plays a key role in conveying information. Similarly, realistic co-speech gestures are crucial to enable natural and smooth interactions with social agents. Current end-to-end co-speech gesture generation systems use a single modality for representing speech: either audio or text. These systems are therefore confined to producing either acoustically-linked beat gestures or semantically-linked gesticulation (e.g., raising a hand when saying "high"): they cannot appropriately learn to generate both gesture types. We present a model designed to produce arbitrary beat and semantic gestures together. Our deep-learning based model takes both acoustic and semantic representations of speech as input, and generates gestures as a sequence of joint angle rotations as output. The resulting gestures can be applied to both virtual agents and humanoid robots. Subjective and objective evaluations confirm the success of our approach. The code and video are available at the project page https://svito-zar.github.io/gesticulator .

Seyyed Saeed Sarfjoo, Xin Wang, Gustav Eje Henter et al.

Nowadays vast amounts of speech data are recorded from low-quality recorder devices such as smartphones, tablets, laptops, and medium-quality microphones. The objective of this research was to study the automatic generation of high-quality speech from such low-quality device-recorded speech, which could then be applied to many speech-generation tasks. In this paper, we first introduce our new device-recorded speech dataset then propose an improved end-to-end method for automatically transforming the low-quality device-recorded speech into professional high-quality speech. Our method is an extension of a generative adversarial network (GAN)-based speech enhancement model called speech enhancement GAN (SEGAN), and we present two modifications to make model training more robust and stable. Finally, from a large-scale listening test, we show that our method can significantly enhance the quality of device-recorded speech signals.

Gustav Eje Henter, Simon Alexanderson, Jonas Beskow

Data-driven modelling and synthesis of motion is an active research area with applications that include animation, games, and social robotics. This paper introduces a new class of probabilistic, generative, and controllable motion-data models based on normalising flows. Models of this kind can describe highly complex distributions, yet can be trained efficiently using exact maximum likelihood, unlike GANs or VAEs. Our proposed model is autoregressive and uses LSTMs to enable arbitrarily long time-dependencies. Importantly, is is also causal, meaning that each pose in the output sequence is generated without access to poses or control inputs from future time steps; this absence of algorithmic latency is important for interactive applications with real-time motion control. The approach can in principle be applied to any type of motion since it does not make restrictive, task-specific assumptions regarding the motion or the character morphology. We evaluate the models on motion-capture datasets of human and quadruped locomotion. Objective and subjective results show that randomly-sampled motion from the proposed method outperforms task-agnostic baselines and attains a motion quality close to recorded motion capture.

Taras Kucherenko, Dai Hasegawa, Gustav Eje Henter et al.

This paper presents a novel framework for automatic speech-driven gesture generation, applicable to human-agent interaction including both virtual agents and robots. Specifically, we extend recent deep-learning-based, data-driven methods for speech-driven gesture generation by incorporating representation learning. Our model takes speech as input and produces gestures as output, in the form of a sequence of 3D coordinates. Our approach consists of two steps. First, we learn a lower-dimensional representation of human motion using a denoising autoencoder neural network, consisting of a motion encoder MotionE and a motion decoder MotionD. The learned representation preserves the most important aspects of the human pose variation while removing less relevant variation. Second, we train a novel encoder network SpeechE to map from speech to a corresponding motion representation with reduced dimensionality. At test time, the speech encoder and the motion decoder networks are combined: SpeechE predicts motion representations based on a given speech signal and MotionD then decodes these representations to produce motion sequences. We evaluate different representation sizes in order to find the most effective dimensionality for the representation. We also evaluate the effects of using different speech features as input to the model. We find that mel-frequency cepstral coefficients (MFCCs), alone or combined with prosodic features, perform the best. The results of a subsequent user study confirm the benefits of the representation learning.

Gustav Eje Henter, Jaime Lorenzo-Trueba, Xin Wang et al.

Generating versatile and appropriate synthetic speech requires control over the output expression separate from the spoken text. Important non-textual speech variation is seldom annotated, in which case output control must be learned in an unsupervised fashion. In this paper, we perform an in-depth study of methods for unsupervised learning of control in statistical speech synthesis. For example, we show that popular unsupervised training heuristics can be interpreted as variational inference in certain autoencoder models. We additionally connect these models to VQ-VAEs, another, recently-proposed class of deep variational autoencoders, which we show can be derived from a very similar mathematical argument. The implications of these new probabilistic interpretations are discussed. We illustrate the utility of the various approaches with an application to acoustic modelling for emotional speech synthesis, where the unsupervised methods for learning expression control (without access to emotional labels) are found to give results that in many aspects match or surpass the previous best supervised approach.

Gustav Eje Henter, Arne Leijon, W. Bastiaan Kleijn

We consider Markov models of stochastic processes where the next-step conditional distribution is defined by a kernel density estimator (KDE), similar to Markov forecast densities and certain time-series bootstrap schemes. The KDE Markov models (KDE-MMs) we discuss are nonlinear, nonparametric, fully probabilistic representations of stationary processes, based on techniques with strong asymptotic consistency properties. The models generate new data by concatenating points from the training data sequences in a context-sensitive manner, together with some additive driving noise. We present novel EM-type maximum-likelihood algorithms for data-driven bandwidth selection in KDE-MMs. Additionally, we augment the KDE-MMs with a hidden state, yielding a new model class, KDE-HMMs. The added state variable captures non-Markovian long memory and signal structure (e.g., slow oscillations), complementing the short-range dependences described by the Markov process. The resulting joint Markov and hidden-Markov structure is appealing for modelling complex real-world processes such as speech signals. We present guaranteed-ascent EM-update equations for model parameters in the case of Gaussian kernels, as well as relaxed update formulas that greatly accelerate training in practice. Experiments demonstrate increased held-out set probability for KDE-HMMs on several challenging natural and synthetic data series, compared to traditional techniques such as autoregressive models, HMMs, and their combinations.

Gustav Eje Henter, Simon King, Thomas Merritt et al.

Output from statistical parametric speech synthesis (SPSS) remains noticeably worse than natural speech recordings in terms of quality, naturalness, speaker similarity, and intelligibility in noise. There are many hypotheses regarding the origins of these shortcomings, but these hypotheses are often kept vague and presented without empirical evidence that could confirm and quantify how a specific shortcoming contributes to imperfections in the synthesised speech. Throughout speech synthesis literature, surprisingly little work is dedicated towards identifying the perceptually most important problems in speech synthesis, even though such knowledge would be of great value for creating better SPSS systems. In this book chapter, we analyse some of the shortcomings of SPSS. In particular, we discuss issues with vocoding and present a general methodology for quantifying the effect of any of the many assumptions and design choices that hold SPSS back. The methodology is accompanied by an example that carefully measures and compares the severity of perceptual limitations imposed by vocoding as well as other factors such as the statistical model and its use.

Sang Phan, Gustav Eje Henter, Yusuke Miyao et al.

Captioning models are typically trained using the cross-entropy loss. However, their performance is evaluated on other metrics designed to better correlate with human assessments. Recently, it has been shown that reinforcement learning (RL) can directly optimize these metrics in tasks such as captioning. However, this is computationally costly and requires specifying a baseline reward at each step to make training converge. We propose a fast approach to optimize one's objective of interest through the REINFORCE algorithm. First we show that, by replacing model samples with ground-truth sentences, RL training can be seen as a form of weighted cross-entropy loss, giving a fast, RL-based pre-training algorithm. Second, we propose to use the consensus among ground-truth captions of the same video as the baseline reward. This can be computed very efficiently. We call the complete proposal Consensus-based Sequence Training (CST). Applied to the MSRVTT video captioning benchmark, our proposals train significantly faster than comparable methods and establish a new state-of-the-art on the task, improving the CIDEr score from 47.3 to 54.2.

Srikanth Ronanki, Oliver Watts, Simon King et al.

This paper proposes a new approach to duration modelling for statistical parametric speech synthesis in which a recurrent statistical model is trained to output a phone transition probability at each timestep (acoustic frame). Unlike conventional approaches to duration modelling -- which assume that duration distributions have a particular form (e.g., a Gaussian) and use the mean of that distribution for synthesis -- our approach can in principle model any distribution supported on the non-negative integers. Generation from this model can be performed in many ways; here we consider output generation based on the median predicted duration. The median is more typical (more probable) than the conventional mean duration, is robust to training-data irregularities, and enables incremental generation. Furthermore, a frame-level approach to duration prediction is consistent with a longer-term goal of modelling durations and acoustic features together. Results indicate that the proposed method is competitive with baseline approaches in approximating the median duration of held-out natural speech.