Evaluation of text-to-speech (TTS) models is currently dominated by Mean-Opinion-Score (MOS) listening test, but MOS has been increasingly questioned for its validity. MOS tests place listeners in a passive setup, in which they do not actively interact with the TTS and usually evaluate isolated utterances without context. Thus it gives no indication for how well a TTS model suits an interactive application like spoken dialogue system, in which the capability of generating appropriate speech in the dialogue context is paramount. We aim to take a first step towards addressing this shortcoming by evaluating several state-of-the-art neural TTS models, including one that adapts to dialogue context, in a custom-built spoken dialogue system. We present system design, experiment setup, and results. Our work is the first to evaluate TTS in contextual dialogue system interactions. We also discuss the shortcomings and future opportunities of the proposed evaluation paradigm.

Recent advances in generative language modeling applied to discrete speech tokens presented a new avenue for text-to-speech (TTS) synthesis. These speech language models (SLMs), similarly to their textual counterparts, are scalable, probabilistic, and context-aware. While they can produce diverse and natural outputs, they sometimes face issues such as unintelligibility and the inclusion of non-speech noises or hallucination. As the adoption of this innovative paradigm in speech synthesis increases, there is a clear need for an in-depth evaluation of its capabilities and limitations. In this paper, we evaluate TTS from a discrete token-based SLM, through both automatic metrics and listening tests. We examine five key dimensions: speaking style, intelligibility, speaker consistency, prosodic variation, spontaneous behaviour. Our results highlight the model's strength in generating varied prosody and spontaneous outputs. It is also rated higher in naturalness and context appropriateness in listening tests compared to a conventional TTS. However, the model's performance in intelligibility and speaker consistency lags behind traditional TTS. Additionally, we show that increasing the scale of SLMs offers a modest boost in robustness. Our findings aim to serve as a benchmark for future advancements in generative SLMs for speech synthesis.

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

Turn-taking is a fundamental aspect of human communication where speakers convey their intention to either hold, or yield, their turn through prosodic cues. Using the recently proposed Voice Activity Projection model, we propose an automatic evaluation approach to measure these aspects for conversational speech synthesis. We investigate the ability of three commercial, and two open-source, Text-To-Speech (TTS) systems ability to generate turn-taking cues over simulated turns. By varying the stimuli, or controlling the prosody, we analyze the models performances. We show that while commercial TTS largely provide appropriate cues, they often produce ambiguous signals, and that further improvements are possible. TTS, trained on read or spontaneous speech, produce strong turn-hold but weak turn-yield cues. We argue that this approach, that focus on functional aspects of interaction, provides a useful addition to other important speech metrics, such as intelligibility and naturalness.

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.

Many robots use their voice to communicate with people in spoken language but the voices commonly used for robots are often optimized for transactional interactions, rather than social ones. This can limit their ability to create engaging and natural interactions. To address this issue, we designed a spontaneous text-to-speech tool and used it to author natural and spontaneous robot speech. A crowdsourcing evaluation methodology is proposed to compare this type of speech to natural speech and state-of-the-art text-to-speech technology, both in disembodied and embodied form. We created speech samples in a naturalistic setting of people playing tabletop games and conducted a user study evaluating Naturalness, Intelligibility, Social Impression, Prosody, and Perceived Intelligence. The speech samples were chosen to represent three contexts that are common in tabletopgames and the contexts were introduced to the participants that evaluated the speech samples. The study results show that the proposed evaluation methodology allowed for a robust analysis that successfully compared the different conditions. Moreover, the spontaneous voice met our target design goal of being perceived as more natural than a leading commercial text-to-speech.

One of the main goals of robotics and intelligent agent research is to enable them to communicate with humans in physically situated settings. Human communication consists of both verbal and non-verbal modes. Recent studies in enabling communication for intelligent agents have focused on verbal modes, i.e., language and speech. However, in a situated setting the non-verbal mode is crucial for an agent to adapt flexible communication strategies. In this work, we focus on learning to generate non-verbal communicative expressions in situated embodied interactive agents. Specifically, we show that an agent can learn pointing gestures in a physically simulated environment through a combination of imitation and reinforcement learning that achieves high motion naturalness and high referential accuracy. We compared our proposed system against several baselines in both subjective and objective evaluations. The subjective evaluation is done in a virtual reality setting where an embodied referential game is played between the user and the agent in a shared 3D space, a setup that fully assesses the communicative capabilities of the generated gestures. The evaluations show that our model achieves a higher level of referential accuracy and motion naturalness compared to a state-of-the-art supervised learning motion synthesis model, showing the promise of our proposed system that combines imitation and reinforcement learning for generating communicative gestures. Additionally, our system is robust in a physically-simulated environment thus has the potential of being applied to robots.

In recent years, numerous speech synthesis systems have been proposed that feature multi-dimensional controllability, generating a level of variability that surpasses traditional TTS systems by orders of magnitude. However, it remains challenging for developers to comprehend and demonstrate the potential of these advanced systems. We introduce So-to-Speak, a customisable interface tailored for showcasing the capabilities of different controllable TTS systems. The interface allows for the generation, synthesis, and playback of hundreds of samples simultaneously, displayed on an interactive grid, with variation both low level prosodic features and high level style controls. To offer insights into speech quality, automatic estimates of MOS scores are presented for each sample. So-to-Speak facilitates the audiovisual exploration of the interaction between various speech features, which can be useful in a range of applications in speech technology.

Inserting fillers (such as "um", "like") to clean speech text has a rich history of study. One major application is to make dialogue systems sound more spontaneous. The ambiguity of filler occurrence and inter-speaker difference make both modeling and evaluation difficult. In this paper, we study sampling-based filler insertion, a simple yet unexplored approach to inserting fillers. We propose an objective score called Filler Perplexity (FPP). We build three models trained on two single-speaker spontaneous corpora, and evaluate them with FPP and perceptual tests. We implement two innovations in perceptual tests, (1) evaluating filler insertion on dialogue systems output, (2) synthesizing speech with neural spontaneous TTS engines. FPP proves to be useful in analysis but does not correlate well with perceptual MOS. Perceptual results show little difference between compared filler insertion models including with ground-truth, which may be due to the ambiguity of what is good filler insertion and a strong neural spontaneous TTS that produces natural speech irrespective of input. Results also show preference for filler-inserted speech synthesized with spontaneous TTS. The same test using TTS based on read speech obtains the opposite results, which shows the importance of using spontaneous TTS in evaluating filler insertions. Audio samples: www.speech.kth.se/tts- demos/LREC22