LoRP-TTS: Low-Rank Personalized Text-To-Speech

TL;DR

LoRP addresses the challenge of creating truly diverse speech corpora by enabling high-fidelity, speaker-specific TTS from single, noisy prompts. It introduces Low-Rank Personalization (LoRP), which applies Low-Rank Adaptation ($r=16$, $\alpha=16$) to Voicebox, adding about $10^7$ parameters (≈$2.3\%$ of weights) and requiring $100$ optimizer steps per prompt. Across multilingual pretraining plus Polish fine-tuning, LoRP delivers up to $30pp$ gains in speaker similarity while maintaining content and naturalness, with strong generalization across Clarin, Fleurs, Nemo, and Kretes. The approach reduces data collection costs and points toward cross-lingual and expressive prosody enhancements, offering practical impact for diverse speech applications.

Abstract

Speech synthesis models convert written text into natural-sounding audio. While earlier models were limited to a single speaker, recent advancements have led to the development of zero-shot systems that generate realistic speech from a wide range of speakers using their voices as additional prompts. However, they still struggle with imitating non-studio-quality samples that differ significantly from the training datasets. In this work, we demonstrate that utilizing Low-Rank Adaptation (LoRA) allows us to successfully use even single recordings of spontaneous speech in noisy environments as prompts. This approach enhances speaker similarity by up to $30pp$ while preserving content and naturalness. It represents a significant step toward creating truly diverse speech corpora, that is crucial in all speech-related tasks.

Figures (3)

• Figure 1: Evaluation metrics for various numbers of samples and optimizer steps on Kretes dataset.
• Figure 2: Evaluation metrics across different datasets.
• Figure 3: Synthesis training pipeline.