Recent work shows promising results in expanding the capabilities of large language models (LLM) to directly understand and synthesize speech. However, an LLM-based strategy for modeling spoken dialogs remains elusive, calling for further investigation. This paper introduces an extensive speech-text LLM framework, the Unified Spoken Dialog Model (USDM), designed to generate coherent spoken responses with naturally occurring prosodic features relevant to the given input speech without relying on explicit automatic speech recognition (ASR) or text-to-speech (TTS) systems. We have verified the inclusion of prosody in speech tokens that predominantly contain semantic information and have used this foundation to construct a prosody-infused speech-text model. Additionally, we propose a generalized speech-text pretraining scheme that enhances the capture of cross-modal semantics. To construct USDM, we fine-tune our speech-text model on spoken dialog data using a multi-step spoken dialog template that stimulates the chain-of-reasoning capabilities exhibited by the underlying LLM. Automatic and human evaluations on the DailyTalk dataset demonstrate that our approach effectively generates natural-sounding spoken responses, surpassing previous and cascaded baselines. Our code and checkpoints are available at https://github.com/naver-ai/usdm.

Recent advances in synthetic speech quality have enabled us to train text-to-speech (TTS) systems by using synthetic corpora. However, merely increasing the amount of synthetic data is not always advantageous for improving training efficiency. Our aim in this study is to selectively choose synthetic data that are beneficial to the training process. In the proposed method, we first adopt a variational autoencoder whose posterior distribution is utilized to extract latent features representing acoustic similarity between the recorded and synthetic corpora. By using those learned features, we then train a ranking support vector machine (RankSVM) that is well known for effectively ranking relative attributes among binary classes. By setting the recorded and synthetic ones as two opposite classes, RankSVM is used to determine how the synthesized speech is acoustically similar to the recorded data. Then, synthetic TTS data, whose distribution is close to the recorded data, are selected from large-scale synthetic corpora. By using these data for retraining the TTS model, the synthetic quality can be significantly improved. Objective and subjective evaluation results show the superiority of the proposed method over the conventional methods.

This paper proposes an effective emotional text-to-speech (TTS) system with a pre-trained language model (LM)-based emotion prediction method.Unlike conventional systems that require auxiliary inputs such as manually defined emotion classes, our system directly estimates emotion-related attributes from the input text.Specifically, we utilize generative pre-trained transformer (GPT)-3 to jointly predict both an emotion class and its strength in representing emotions' coarse and fine properties, respectively.Then, these attributes are combined in the emotional embedding space and used as conditional features of the TTS model for generating output speech signals.Consequently, the proposed system can produce emotional speech only from text without any auxiliary inputs.Furthermore, because the GPT-3 enables to capture emotional context among the consecutive sentences, the proposed method can effectively handle the paragraph-level generation of emotional speech.

This paper proposes a spectral-domain perceptual weighting technique for Parallel WaveGAN-based text-to-speech (TTS) systems. The recently proposed Parallel WaveGAN vocoder successfully generates waveform sequences using a fast non-autoregressive WaveNet model. By employing multi-resolution short-time Fourier transform (MR-STFT) criteria with a generative adversarial network, the light-weight convolutional networks can be effectively trained without any distillation process. To further improve the vocoding performance, we propose the application of frequency-dependent weighting to the MR-STFT loss function. The proposed method penalizes perceptually-sensitive errors in the frequency domain; thus, the model is optimized toward reducing auditory noise in the synthesized speech. Subjective listening test results demonstrate that our proposed method achieves 4.21 and 4.26 TTS mean opinion scores for female and male Korean speakers, respectively.

This paper proposes a modeling-by-generation (MbG) excitation vocoder for a neural text-to-speech (TTS) system. Recently proposed neural excitation vocoders can realize qualified waveform generation by combining a vocal tract filter with a WaveNet-based glottal excitation generator. However, when these vocoders are used in a TTS system, the quality of synthesized speech is often degraded owing to a mismatch between training and synthesis steps. Specifically, the vocoder is separately trained from an acoustic model front-end. Therefore, estimation errors of the acoustic model are inevitably boosted throughout the synthesis process of the vocoder back-end. To address this problem, we propose to incorporate an MbG structure into the vocoder's training process. In the proposed method, the excitation signal is extracted by the acoustic model's generated spectral parameters, and the neural vocoder is then optimized not only to learn the target excitation's distribution but also to compensate for the estimation errors occurring from the acoustic model. Furthermore, as the generated spectral parameters are shared in the training and synthesis steps, their mismatch conditions can be reduced effectively. The experimental results verify that the proposed system provides high-quality synthetic speech by achieving a mean opinion score of 4.57 within the TTS framework.

In this letter, we propose a high-quality emotional speech synthesis system, using emotional vector space, i.e., the weighted sum of global style tokens (GSTs). Our previous research verified the feasibility of GST-based emotional speech synthesis in an end-to-end text-to-speech synthesis framework. However, selecting appropriate reference audio (RA) signals to extract emotion embedding vectors to the specific types of target emotions remains problematic. To ameliorate the selection problem, we propose an effective way of generating emotion embedding vectors by utilizing the trained GSTs. By assuming that the trained GSTs represent an emotional vector space, we first investigate the distribution of all the training samples depending on the type of each emotion. We then regard the centroid of the distribution as an emotion-specific weighting value, which effectively controls the expressiveness of synthesized speech, even without using the RA for guidance, as it did before. Finally, we confirm that the proposed controlled weight-based method is superior to the conventional emotion label-based methods in terms of perceptual quality and emotion classification accuracy.