Wavehax: Aliasing-Free Neural Waveform Synthesis Based on 2D Convolution and Harmonic Prior for Reliable Complex Spectrogram Estimation (2411.06807v1)

Abstract: Neural vocoders often struggle with aliasing in latent feature spaces, caused by time-domain nonlinear operations and resampling layers. Aliasing folds high-frequency components into the low-frequency range, making aliased and original frequency components indistinguishable and introducing two practical issues. First, aliasing complicates the waveform generation process, as the subsequent layers must address these aliasing effects, increasing the computational complexity. Second, it limits extrapolation performance, particularly in handling high fundamental frequencies, which degrades the perceptual quality of generated speech waveforms. This paper demonstrates that 1) time-domain nonlinear operations inevitably introduce aliasing but provide a strong inductive bias for harmonic generation, and 2) time-frequency-domain processing can achieve aliasing-free waveform synthesis but lacks the inductive bias for effective harmonic generation. Building on this insight, we propose Wavehax, an aliasing-free neural WAVEform generator that integrates 2D convolution and a HArmonic prior for reliable Complex Spectrogram estimation. Experimental results show that Wavehax achieves speech quality comparable to existing high-fidelity neural vocoders and exhibits exceptional robustness in scenarios requiring high fundamental frequency extrapolation, where aliasing effects become typically severe. Moreover, Wavehax requires less than 5% of the multiply-accumulate operations and model parameters compared to HiFi-GAN V1, while achieving over four times faster CPU inference speed.

• The paper introduces Wavehax, an aliasing-free neural vocoder using 2D CNNs and harmonic priors for robust complex spectrogram estimation.
• Wavehax demonstrates significant computational efficiency, requiring less than 5% of the operations of HiFi-GAN V1 and achieving over 4x faster CPU inference.
• This novel approach offers high-fidelity synthesis even with challenging high F0s and is well-suited for low-resource environments and broader audio synthesis applications.

Overview of Wavehax: Aliasing-Free Neural Waveform Synthesis

The paper "Wavehax: Aliasing-Free Neural Waveform Synthesis" presents Wavehax, a neural vocoder designed to eliminate aliasing artifacts commonly encountered in neural waveform synthesis. This research leverages both time-frequency domain processing and harmonic priors to produce high-fidelity and robust speech synthesis while significantly reducing computational demands compared to existing models like HiFi-GAN V1.

Background and Motivations

Neural vocoders, integral to speech synthesis, transcode acoustic features into audio waveforms. Traditional models primarily focus on time-domain processing, employing autoregressive and GAN-based architectures. These approaches are susceptible to aliasing, a phenomenon where high-frequency components fold back into lower frequencies. Aliasing introduces distortion and complicates waveform synthesis, especially in high-frequency scenarios. Recent studies have looked into countermeasures such as temporal upsampling and architectural modifications in GANs to alleviate aliasing, yet completely eliminating it remains challenging.

Contributions of Wavehax

Wavehax innovatively integrates 2D convolutional networks with harmonic priors for complex spectrogram estimation, advancing the field in several key areas:

1. Aliasing Mitigation: By operating primarily in the time-frequency domain, Wavehax avoids the aliasing associated with time-domain nonlinear operations. This approach provides a solid theoretical and practical framework for aliasing-free synthesis.
2. Computational Efficiency: Wavehax achieves considerable reductions in computational complexity and model size. Notably, it requires less than 5% of the operations and parameters of HiFi-GAN while delivering more than four times faster CPU inference speed.
3. High-Fidelity Output: Despite its reduced complexity, Wavehax maintains or surpasses the speech quality of high-fidelity vocoders, proving particularly robust in handling high fundamental frequency (F0) extrapolations where aliasing issues are pronounced.

Technical Implementation

Wavehax introduces the utilization of 2D CNNs for capturing and processing complex spectrograms, which effectively accommodates the harmonic characteristics of speech. The model leverages the short-time Fourier transform (STFT) to manipulate spectrograms while aligning them with acoustic features extracted from the input waveforms. Harmonic priors imbue the model with inductive biases that facilitate the accurate generation of harmonic components—a crucial aspect for natural-sounding speech synthesis. These innovations collectively enhance the synthesis process's stability and robustness.

Implications and Future Directions

Wavehax opens several avenues for future research and practical application:

• Robustness Across Domains: The integration of harmonic priors and STFT positions Wavehax as a versatile model adaptable to various audio synthesis tasks beyond speech, such as music generation, where harmonic structures play a critical role.
• Optimizing Low-Resource Environments: The drastic reduction in computational demands makes Wavehax suitable for deployment in resource-constrained environments or applications requiring real-time processing, such as virtual assistants.
• Improvement of Vocoder Architectures: By demonstrating the effectiveness of aliasing-free synthesis, the paper paves the way for refining vocoder architectures, fostering further innovations that may blend Wavehax's principles with emerging models like diffusion probabilistic vocoders.

In summary, Wavehax presents a significant step towards high-efficiency, aliasing-free neural vocoders, aligning theoretical insights with practical advancements. The model not only enhances speech synthesis quality and efficiency but also offers a robust framework adaptable to diverse audio synthesis domains. Future work might explore further optimization, real-time applications, and extension to other non-speech audio domains, potentially broadening the impact of this research across the field of audio processing.