Non-intrusive Speech Quality Assessment with Diffusion Models Trained on Clean Speech

2024-10-23

arXiv.org

AI Summary

This paper addresses no-reference voice quality assessment (NR-VQA). We propose a non-intrusive method based on unconditional diffusion models, trained exclusively on clean speech—without distorted samples, reference utterances, or human annotations. The approach employs a deterministic noise scheduling scheme to map speech signals to a Gaussian latent space; the log-likelihood at the terminal diffusion step serves as a distortion-sensitive quality score. Key contributions include: (i) the first application of diffusion models to density-estimation-based VQA; (ii) a fully unsupervised, reference-free framework; and (iii) implicit modeling of quality degradation via the clean-speech prior. Experiments demonstrate strong correlation with reference-based metrics (|r| > 0.92 vs. POLQA and SI-SDR) across diverse distortion types, and substantial gains over state-of-the-art NR-VQA methods.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Speech
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Diffusion models have found great success in generating high quality, natural samples of speech, but their potential for density estimation for speech has so far remained largely unexplored. In this work, we leverage an unconditional diffusion model trained only on clean speech for the assessment of speech quality. We show that the quality of a speech utterance can be assessed by estimating the likelihood of a corresponding sample in the terminating Gaussian distribution, obtained via a deterministic noising process. The resulting method is purely unsupervised, trained only on clean speech, and therefore does not rely on annotations. Our diffusion-based approach leverages clean speech priors to assess quality based on how the input relates to the learned distribution of clean data. Our proposed log-likelihoods show promising results, correlating well with intrusive speech quality metrics such as POLQA and SI-SDR.

Problem

• Assess speech quality non-intrusively using diffusion models
• Estimate likelihood in Gaussian distribution for quality evaluation
• Leverage clean speech priors without relying on annotations

Innovation

• Unsupervised diffusion model for speech quality
• Clean speech priors for quality assessment
• Log-likelihood correlates with human scores