Differentiable Time-Varying Linear Prediction in the Context of End-to-End Analysis-by-Synthesis

TL;DR

This paper introduces an efficient, differentiable, sample-wise time-varying linear prediction method for audio synthesis, improving end-to-end training and voice reconstruction quality over existing frame-wise approaches.

Contribution

It generalizes the GOLF vocoder's LP implementation to time-varying cases, enabling better end-to-end training and higher quality voice synthesis.

Findings

GOLF with time-varying LP outperforms frame-wise versions in voice reconstruction.

Synthesized voices from GOLF scored higher than state-of-the-art WORLD vocoder.

The method enables faster, more accurate end-to-end audio synthesis.

Abstract

Training the linear prediction (LP) operator end-to-end for audio synthesis in modern deep learning frameworks is slow due to its recursive formulation. In addition, frame-wise approximation as an acceleration method cannot generalise well to test time conditions where the LP is computed sample-wise. Efficient differentiable sample-wise LP for end-to-end training is the key to removing this barrier. We generalise the efficient time-invariant LP implementation from the GOLF vocoder to time-varying cases. Combining this with the classic source-filter model, we show that the improved GOLF learns LP coefficients and reconstructs the voice better than its frame-wise counterparts. Moreover, in our listening test, synthesised outputs from GOLF scored higher in quality ratings than the state-of-the-art differentiable WORLD vocoder.