Speech is one of the most fundamental means of human communication and plays an indispensable role in our daily lives. In tasks such as audiobook narration or voice-based interaction with AI systems, collecting large-scale speech data from human speak...
Speech is one of the most fundamental means of human communication and plays an indispensable role in our daily lives. In tasks such as audiobook narration or voice-based interaction with AI systems, collecting large-scale speech data from human speakers is often impractical and costly. As a result, deep learning-based speech synthesis has emerged as a viable and scalable solution. While the core objective is to generate natural and high-quality speech, the specific characteristics required vary by application—for example, capturing emotion, mimicking a particular speaker’s timbre, or producing conversational and affective speech that resembles real-world dialog. Among these, speaker-adaptive speech synthesis, also referred to as personalized speech synthesis, has received considerable attention due to its diverse applications.
This dissertation presents a comprehensive methodology and analysis of personalized speech synthesis across multiple tasks, including text-to-speech (TTS), voice conversion (VC), and spoken dialog modeling. We first introduce a unified framework for building personalized TTS and VC systems, which requires only 5–10 seconds of untranscribed speech by fine-tuning a multi-speaker model. We then enhance the framework’s parameter efficiency by identifying key parameters for speaker adaptation and exploring various strategies to strengthen speaker conditioning. Finally, we propose a spoken dialog model capable of generating affective speech responses in a target speaker's voice, guided by a short reference audio sample.
To construct a unified framework for personalized TTS and VC, we propose UnitSpeech, a fine-tuning-based, speaker-adaptive speech synthesis model that enables adaptation from minimal untranscribed speech by replacing textual input with phonetic units. These units, a type of semantic token, are self-supervised speech representations known to capture the linguistic content of speech. In this framework, a unit encoder replaces the conventional text encoder to process these semantic tokens, thereby eliminating the need for transcripts during model fine-tuning. The personalized decoder, fine-tuned on unit-speech pairs, supports both speaker-adaptive TTS through integration with the text encoder and any-to-any VC via the unit encoder. UnitSpeech achieves performance comparable to or surpassing strong baselines and demonstrates notable robustness on real-world data.
To reduce the storage required per speaker in the fine-tuning-based, speaker-adaptive speech synthesis model, we further propose VoiceTailor, a model with parameter-efficient fine-tuning strategy that identifies and adapts only key parameters, specifically the linear layers in attention modules, through low-rank adapters. In addition, to achieve strong speaker adaptation performance with a minimal number of trainable parameters, we explore various guiding strategies that enhance speaker information during speech synthesis, resulting in optimal performance. This approach reduces the fine-tuned parameter count to just 0.25% of the full model, while maintaining comparable speaker similarity and audio quality.
To extend speaker adaptation capabilities to a speech large language model (speech LLM) for spoken dialog, we introduce the Unified Spoken Dialog Model (USDM). This end-to-end model generates coherent and prosodically natural responses in the voice of a target speaker, without relying on explicit automatic speech recognition (ASR) or TTS modules. Initially, we demonstrate that our selected speech tokens preserve both semantic content and detailed prosodic features. We then modify a well-established personalized speech decoder framework and incorporate it into our speech LLM. Within this integrated model, in conjunction with the proposed speech LLM pretraining method, our model generates expressive and personalized spoken responses.