Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GOMPSNR: Reflourish the Signal-to-Noise Ratio Metric for Audio Generation Tasks

Lingling Dai, Andong Li, Cheng Chi, Yifan Liang, Xiaodong Li, Chengshi Zheng

TL;DR

This work tackles the misalignment between SNR-based metrics and human perception in audio generation. It introduces GOMPSNR, a phase-aware metric that replaces the traditional instantaneous phase with omnidirectional phase derivatives $\nabla_i\theta$ and includes a corrected correlation term to improve stability and perceptual relevance. It further proposes phase-oriented losses, including magnitude-guided refinement (WOP) and joint magnitude–phase losses (ORI, CORI), demonstrating that a carefully chosen combination yields substantial gains across multiple vocoders and neural audio codecs. Across LJSpeech, LibriTTS, and auxiliary codec evaluations, GOMPSNR shows stronger correlation with perceptual metrics and the proposed losses yield clear improvements in objective quality and intelligibility.

Abstract

In the field of audio generation, signal-to-noise ratio (SNR) has long served as an objective metric for evaluating audio quality. Nevertheless, recent studies have shown that SNR and its variants are not always highly correlated with human perception, prompting us to raise the questions: Why does SNR fail in measuring audio quality? And how to improve its reliability as an objective metric? In this paper, we identify the inadequate measurement of phase distance as a pivotal factor and propose to reformulate SNR with specially designed phase-distance terms, yielding an improved metric named GOMPSNR. We further extend the newly proposed formulation to derive two novel categories of loss function, corresponding to magnitude-guided phase refinement and joint magnitude-phase optimization, respectively. Besides, extensive experiments are conducted for an optimal combination of different loss functions. Experimental results on advanced neural vocoders demonstrate that our proposed GOMPSNR exhibits more reliable error measurement than SNR. Meanwhile, our proposed loss functions yield substantial improvements in model performance, and our wellchosen combination of different loss functions further optimizes the overall model capability.

GOMPSNR: Reflourish the Signal-to-Noise Ratio Metric for Audio Generation Tasks

TL;DR

This work tackles the misalignment between SNR-based metrics and human perception in audio generation. It introduces GOMPSNR, a phase-aware metric that replaces the traditional instantaneous phase with omnidirectional phase derivatives and includes a corrected correlation term to improve stability and perceptual relevance. It further proposes phase-oriented losses, including magnitude-guided refinement (WOP) and joint magnitude–phase losses (ORI, CORI), demonstrating that a carefully chosen combination yields substantial gains across multiple vocoders and neural audio codecs. Across LJSpeech, LibriTTS, and auxiliary codec evaluations, GOMPSNR shows stronger correlation with perceptual metrics and the proposed losses yield clear improvements in objective quality and intelligibility.

Abstract

In the field of audio generation, signal-to-noise ratio (SNR) has long served as an objective metric for evaluating audio quality. Nevertheless, recent studies have shown that SNR and its variants are not always highly correlated with human perception, prompting us to raise the questions: Why does SNR fail in measuring audio quality? And how to improve its reliability as an objective metric? In this paper, we identify the inadequate measurement of phase distance as a pivotal factor and propose to reformulate SNR with specially designed phase-distance terms, yielding an improved metric named GOMPSNR. We further extend the newly proposed formulation to derive two novel categories of loss function, corresponding to magnitude-guided phase refinement and joint magnitude-phase optimization, respectively. Besides, extensive experiments are conducted for an optimal combination of different loss functions. Experimental results on advanced neural vocoders demonstrate that our proposed GOMPSNR exhibits more reliable error measurement than SNR. Meanwhile, our proposed loss functions yield substantial improvements in model performance, and our wellchosen combination of different loss functions further optimizes the overall model capability.
Paper Structure (19 sections, 8 equations, 5 figures, 4 tables)

This paper contains 19 sections, 8 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Audio Generation Tasks
  4. Objective Metrics for Audio Quality Assessment
  5. Loss Functions for Enhancing the Audio Quality
  6. Methodology
  7. Omnidirectional Phase Derivatives
  8. Generalized Omnidirectional Phase-oriented SNR
  9. Coupling Magnitude and Phase in Loss Functions
  10. Implementation Details
  11. Data Preparation
  12. Training Settings
  13. Results and Discussion
  14. Validation on GOMPSNR
  15. Phase-oriented Loss Functions
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Illustration of the performance improvement brought by our proposed methods. (a) The correlation of SNR and our proposed GOMPSNR with other perceptual metrics. (b) The performance comparison between the original loss and our proposed loss.
  • Figure 2: Illustration of obtaining the omnidirectional phase derivatives.
  • Figure 3: The visualization of the spectrogram discrepancy in terms of (a) magnitude and (b) phase.
  • Figure 4: The correlation of SNR and GOMPSNR with several commly adopted perceptual metrics in terms of PCC and SRCC on the LibriTTS dataset.
  • Figure 5: The objective results of different codec models on the LibriTTS Dataset.