AF-Vocoder: Artifact-Free Neural Vocoder with Global Artifact Filter

Zhuangqi Chen, Xianjun Xia, Xiaohuai Le, Siyu Sun, Chuanzeng HuangInterspeech 2025

TL;DR

This paper introduces AF-Vocoder, a novel GAN-based neural vocoder that significantly reduces audible artifacts like aliasing and blurring in synthesized speech. It achieves this by incorporating a new frequency-domain "Global Artifact Filter" (GAFilter) that learns to suppress unwanted frequency components, leading to higher fidelity and more stable training.

What Problem Does This Paper Address?

Synthesizing realistic human speech from acoustic features (like mel-spectrograms) is a crucial task in areas like text-to-speech and voice conversion. Generative Adversarial Network (GAN)-based vocoders have shown great promise, offering high-fidelity synthesis and fast inference. However, a persistent challenge is the generation of unwanted artifacts, primarily:

  • Aliasing artifacts: These arise from imperfect upsampling processes, where high-frequency components are incorrectly replicated.
  • Blurring artifacts: These occur when the model fails to accurately reconstruct the fine spectral details, leading to a loss of clarity and harmonic richness.

While existing methods have attempted to mitigate these issues, they often focus on only one type of artifact or require complex architectural modifications. This paper aims to provide a more unified and effective solution for generating artifact-free speech.

Methods

The core innovation of AF-Vocoder lies in its Global Artifact Filter (GAFilter). This module is designed to operate in the frequency domain and learn to suppress artifacts.

  • GAFilter Architecture: It consists of a Short-Time Fourier Transform (STFT) module to convert features into the frequency domain, a learnable frequency filter (parameterized by W), and an Inverse STFT (ISTFT) module to reconstruct the time-domain signal. The filter W is designed to enforce a "frequency control inductive bias," learning to identify and remove undesirable frequency components.
  • Integration into Vocoder: GAFilter is integrated into the final upsampling stage of the generator. The authors propose an "Artifact-Free MPC" (AF-MPC) module, which replaces the standard MPC module in the last layer. This placement is strategic: by the final layer, the temporal resolution is fully restored, making it an opportune moment to clean up artifacts before final waveform generation.
  • Overall Generator Architecture: AF-Vocoder builds upon the successful BigVGAN architecture, utilizing stacked ConvTranspose1D layers for upsampling and Multi-Periodicity Composition (MPC) modules with Snake activations for periodic inductive bias. A key simplification compared to BigVGAN is the removal of intermediate upsampling-downsampling operations, as GAFilter is deemed sufficient for artifact suppression.

Datasets

  • Training: The full training set of the LibriTTS dataset was used.
  • Evaluation:
    • In-domain: A subset from the dev-other and dev-clean sets of LibriTTS.
  • Out-of-domain (unseen speakers): LJSpeech, VCTK, and MUSDB18-HD datasets were used to test robustness.
  • Benchmark Consistency: The data selection and configuration for LibriTTS evaluation were kept consistent with BigVGAN to ensure fair comparison. All data was resampled to 24kHz.

Key Insights and Ideas

  • Frequency Domain Filtering for Artifacts: The paper demonstrates that operating in the frequency domain with a learned filter is a powerful way to address both aliasing and blurring artifacts simultaneously.
  • GAFilter as a General Module: Applying GAFilter to an existing vocoder (BigVGAN+GAFilter) showed significant improvements, suggesting its potential as a plug-and-play module for other GAN-based vocoders.
  • Simplification through Effective Filtering: By introducing GAFilter, the authors were able to remove intermediate upsampling-downsampling stages present in prior work (like BigVGAN), leading to a structurally simpler generator without sacrificing performance.
  • Improved Training Stability: The proposed AF-Vocoder exhibited more stable gradient behavior during training, particularly in the early stages, which is attributed to the learned frequency control of GAFilter.

Key Results

  • Objective Metrics: AF-Vocoder consistently outperformed other state-of-the-art GAN-based vocoders (including BigVGAN, BigVSAN, Vocos, FreeV, and Vocos) across multiple objective metrics like M-STFT, PESQ, Periodicity error, V/UV F1, and UTMOS on the LibriTTS dataset.
  • Out-of-Domain Performance: AF-Vocoder showed superior performance on unseen speaker scenarios (LJSpeech, VCTK, MUSDB), highlighting its robustness.
  • Model Size: AF-Vocoder achieved these improvements with comparable or even reduced model sizes (e.g., AF-Vocoder V2 and V3) compared to BigVGAN.
  • Qualitative Results: Visualizations (Figure 2) showed that AF-Vocoder generated speech with significantly fewer blurring artifacts and clearer harmonic structures compared to BigVGAN.
  • Training Stability: Gradient curves (Figure 3) indicated that AF-Vocoder training was more stable and less divergent than BigVGAN.

Limitations and Future Work

  • Limitations: The paper notes that while GAFilter significantly reduces artifacts, the SAN training framework (used in BigVSAN) might introduce more audible artifacts, leading them to focus on Least Squares GAN training for further analysis.
  • Future Work: The authors suggest that the GAFilter module's effectiveness in improving performance and ensuring stable training opens avenues for further research into learned frequency control for artifact suppression in neural vocoders.

Summary updated

BibTeX
@inproceedings{chen2025af,
  author = {Zhuangqi Chen and Xianjun Xia and Xiaohuai Le and Siyu Sun and Chuanzeng Huang},
  title = {{AF-Vocoder: Artifact-Free Neural Vocoder with Global Artifact Filter}},
  booktitle = {Interspeech},
  year = {2025},
}
Figure