Investigating self-supervised features for expressive, multilingual voice conversion

Voice conversion (VC) systems are widely used for several applications, from speaker anonymisation to personalised speech synthesis. Supervised approaches learn a mapping between different speakers using parallel data, which is expensive to produce. Unsupervised approaches are typically trained to reconstruct the input signal, which is composed of the content and the speaker information. Disentangling these components is a challenge and often leads to speaker leakage or prosodic information removal. In this paper, we explore voice conversion by leveraging the potential of self-supervised learning (SSL). A combination of the latent representations of SSL models, concatenated with speaker embeddings, is fed to a vocoder which is trained to reconstruct the input. Zero-shot voice conversion results show that this approach allows to keep the prosody and content of the source speaker while matching the speaker similarity of a VC system based on phonetic posteriorgrams (PPGs).

Enhancing the Stability of LLM-based Speech Generation Systems through Self-Supervised Representations

Large Language Models (LLMs) are one of the most promising technologies for the next era of speech generation systems, due to their scalability and in-context learning capabilities. Nevertheless, they suffer from multiple stability issues at inference time, such as hallucinations, content skipping or speech repetitions. In this work, we introduce a new self-supervised Voice Conversion (VC) architecture which can be used to learn to encode transitory features, such as content, separately from stationary ones, such as speaker ID or recording conditions, creating speaker-disentangled representations. Using speaker-disentangled codes to train LLMs for text-to-speech (TTS) allows the LLM to generate the content and the style of the speech only from the text, similarly to humans, while the speaker identity is provided by the decoder of the VC model. Results show that LLMs trained over speaker-disentangled self-supervised representations provide an improvement of 4.7pp in speaker similarity over SOTA entangled representations, and a word error rate (WER) 5.4pp lower. Furthermore, they achieve higher naturalness than human recordings of the LibriTTS test-other dataset. Finally, we show that using explicit reference embedding negatively impacts intelligibility (stability), with WER increasing by 14pp compared to the model that only uses text to infer the style.

Comparing normalizing flows and diffusion models for prosody and acoustic modelling in text-to-speech

Neural text-to-speech systems are often optimized on L1/L2 losses, which make strong assumptions about the distributions of the target data space. Aiming to improve those assumptions, Normalizing Flows and Diffusion Probabilistic Models were recently proposed as alternatives. In this paper, we compare traditional L1/L2-based approaches to diffusion and flow-based approaches for the tasks of prosody and mel-spectrogram prediction for text-to-speech synthesis. We use a prosody model to generate log-f0 and duration features, which are used to condition an acoustic model that generates mel-spectrograms. Experimental results demonstrate that the flow-based model achieves the best performance for spectrogram prediction, improving over equivalent diffusion and L1 models. Meanwhile, both diffusion and flow-based prosody predictors result in significant improvements over a typical L2-trained prosody models.