Automatic Voice Identification after Speech Resynthesis using PPG

TL;DR

This paper investigates whether Phonetic PosteriorGrams ($PPG$) can underpin a speech resynthesis system that preserves linguistic content while concealing the source speaker identity. By training a $PPG2Mel$ pipeline to convert $PPG$ inputs into Mel-spectrograms and subsequently audio via a WaveGlow vocoder, the authors demonstrate comparable audio quality to a text-to-speech baseline, with vocoder quality contributing most to degradation. Crucially, automatic speaker verification remains unable to recover the source speaker after resynthesis, even when models are trained on synthetic data, indicating low leakage of speaker identity through $PPG$-driven resynthesis. These results support using $PPG$ as a controllable, speaker-agnostic representation for speech edition and voice-resynthesis tasks, while highlighting future work on vocoders and perceptual speaker-similarity evaluations. The work advances practical pathways for editor-controlled speech generation and challenges assumptions about speaker leakage in phonetic representations.

Abstract

Speech resynthesis is a generic task for which we want to synthesize audio with another audio as input, which finds applications for media monitors and journalists.Among different tasks addressed by speech resynthesis, voice conversion preserves the linguistic information while modifying the identity of the speaker, and speech edition preserves the identity of the speaker but some words are modified.In both cases, we need to disentangle speaker and phonetic contents in intermediate representations.Phonetic PosteriorGrams (PPG) are a frame-level probabilistic representation of phonemes, and are usually considered speaker-independent.This paper presents a PPG-based speech resynthesis system.A perceptive evaluation assesses that it produces correct audio quality.Then, we demonstrate that an automatic speaker verification model is not able to recover the source speaker after re-synthesis with PPG, even when the model is trained on synthetic data.

Figures (3)

• Figure 1: Overview of our PPG-based re-synthesis (PPG2Mel) approach. Blue box denotes speaker-specific model. N stands for natural speech, while R stands for re-synthesized speech (see Sec. \ref{['sec:odn']})
• Figure 2: Block representation of the 4 variations of each sample of our perceptive evaluation. Blue boxes are speaker-specific models
• Figure 3: PPG example for: "Such risks can be lessened when the President recognizes the security problem"