This paper proposes a multilingual, multi-speaker (MM) TTS system by using a voice conversion (VC)-based data augmentation method. Creating an MM-TTS model is challenging, owing to the difficulties of collecting polyglot data from multiple speakers. To address this problem, we adopt a cross-lingual, multi-speaker VC model trained with multiple speakers’ monolingual databases. As this model effectively transfers acoustic attributes while retaining the content information, it is possible to generate each speaker’s polyglot corpora. Subsequently, we design the MM-TTS model with variational autoencoder (VAE)-based posterior embeddings. It is to be noted that incorporating VC-augmented polyglot corpora into the TTS training process might degrade synthetic quality, since the corpora sometimes contain unwanted artifacts. To mitigate this issue, the VAE is trained to capture the acoustic dissimilarity between the recorded and VC-augmented datasets. Through the selective choice of the posterior embeddings obtained from the original recordings in the training set, the proposed model enables the generation of acoustically clearer voices.

For personalized speech generation, a neural text-to-speech (TTS) model must be successfully implemented with limited data from a target speaker. To this end, the baseline TTS model needs to be amply generalized to out-of-domain data (i.e., target speaker's speech). However, approaches to address this out-of-domain generalization problem in TTS have yet to be thoroughly studied. In this work, we propose an effective pruning method for a transformer known as sparse attention, to improve the TTS model's generalization abilities. In particular, we prune off redundant connections from self-attention layers whose attention weights are below the threshold. To flexibly determine the pruning strength for searching optimal degree of generalization, we also propose a new differentiable pruning method that allows the model to automatically learn the thresholds. Evaluations on zero-shot multi-speaker TTS verify the effectiveness of our method in terms of voice quality and speaker similarity.

This letter proposes an effective speaker-conditioning method that is applicable to zero-shot multi-speaker text-to-speech (ZSM-TTS) systems. Based on the inductive bias in the speech generation task, in which local context information in text/phoneme sequences heavily affect the speaker characteristics of the output speech, we propose a Speaker-Conditional Convolutional Neural Network (SC-CNN) for the ZSM-TTS task. SC-CNN first predicts convolutional kernels from each learned speaker embedding, then applies 1-D convolutions to phoneme sequences with the predicted kernels. It utilizes the aforementioned inductive bias and effectively models the characteristic of speech by providing the speaker-specific local context in phonetic domain. We also build both FastSpeech2 and VITS-based ZSM-TTS systems to verify its superiority over conventional speaker conditioning methods. The results confirm that the models with SC-CNN outperform the recent ZSM-TTS models in terms of both subjective and objective measurements.

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.

Data augmentation via voice conversion (VC) has been successfully applied to low-resource expressive text-to-speech (TTS) when only neutral data for the target speaker are available. Although the quality of VC is crucial for this approach, it is challenging to learn a stable VC model because the amount of data is limited in low-resource scenarios, and highly expressive speech has large acoustic variety. To address this issue, we propose a novel data augmentation method that combines pitch-shifting and VC techniques. Because pitch-shift data augmentation enables the coverage of a variety of pitch dynamics, it greatly stabilizes training for both VC and TTS models, even when only 1,000 utterances of the target speaker's neutral data are available. Subjective test results showed that a FastSpeech 2-based emotional TTS system with the proposed method improved naturalness and emotional similarity compared with 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.

While generative adversarial networks (GANs) based neural text-to-speech (TTS) systems have shown significant improvement in neural speech synthesis, there is no TTS system to learn to synthesize speech from text sequences with only adversarial feedback. Because adversarial feedback alone is not sufficient to train the generator, current models still require the reconstruction loss compared with the ground-truth and the generated mel-spectrogram directly. In this paper, we present Multi-SpectroGAN (MSG), which can train the multi-speaker model with only the adversarial feedback by conditioning a self-supervised hidden representation of the generator to a conditional discriminator. This leads to better guidance for generator training. Moreover, we also propose adversarial style combination (ASC) for better generalization in the unseen speaking style and transcript, which can learn latent representations of the combined style embedding from multiple mel-spectrograms. Trained with ASC and feature matching, the MSG synthesizes a high-diversity mel-spectrogram by controlling and mixing the individual speaking styles (e.g., duration, pitch, and energy). The result shows that the MSG synthesizes a high-fidelity mel-spectrogram, which has almost the same naturalness MOS score as the ground-truth mel-spectrogram.

In recent works, a flow-based neural vocoder has shown significant improvement in real-time speech generation task. The sequence of invertible flow operations allows the model to convert samples from simple distribution to audio samples. However, training a continuous density model on discrete audio data can degrade model performance due to the topological difference between latent and actual distribution. To resolve this problem, we propose audio dequantization methods in flow-based neural vocoder for high fidelity audio generation. Data dequantization is a well-known method in image generation but has not yet been studied in the audio domain. For this reason, we implement various audio dequantization methods in flow-based neural vocoder and investigate the effect on the generated audio. We conduct various objective performance assessments and subjective evaluation to show that audio dequantization can improve audio generation quality. From our experiments, using audio dequantization produces waveform audio with better harmonic structure and fewer digital artifacts.