BitTTS: Highly Compact Text-to-Speech Using 1.58-bit Quantization and Weight Indexing

2025-06-04

AI Summary

Deploying high-quality text-to-speech (TTS) models on resource-constrained edge devices demands simultaneous optimization of synthesis fidelity and model footprint. This paper proposes an ultra-lightweight TTS compression framework featuring two key innovations: (1) the first 1.58-bit quantization-aware training (QAT) for edge TTS—effectively approximating ternary weights {−1, 0, 1}; and (2) Weight Indexing, a novel technique that maps multiple low-bit weight groups to a single int8 index, drastically reducing both memory storage and computational overhead. Evaluated on standard benchmarks, our method achieves 83% model size reduction while preserving speech naturalness and outperforming non-quantized baselines of comparable size. The core contributions are: (i) the first edge-deployable TTS system supporting 1.58-bit QAT; (ii) an efficient int8-indexed weight representation scheme; and (iii) real-time, high-fidelity speech synthesis on severely resource-limited hardware.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Natural Language Processing: Speech
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

This paper proposes a highly compact, lightweight text-to-speech (TTS) model for on-device applications. To reduce the model size, the proposed model introduces two techniques. First, we introduce quantization-aware training (QAT), which quantizes model parameters during training to as low as 1.58-bit. In this case, most of 32-bit model parameters are quantized to ternary values {-1, 0, 1}. Second, we propose a method named weight indexing. In this method, we save a group of 1.58-bit weights as a single int8 index. This allows for efficient storage of model parameters, even on hardware that treats values in units of 8-bit. Experimental results demonstrate that the proposed method achieved 83 % reduction in model size, while outperforming the baseline of similar model size without quantization in synthesis quality.

Problem

• Reducing TTS model size for on-device applications
• Quantizing model parameters to 1.58-bit using QAT
• Efficient storage via weight indexing with int8

Innovation

• Uses 1.58-bit quantization for compactness
• Employs weight indexing for efficient storage
• Quantization-aware training enhances synthesis quality