Total: 1
A phonetic posteriorgram (PPG) is a time-varying categorical distribution over acoustic units of speech (e.g., phonemes). PPGs are a popular representation in speech generation due to their ability to disentangle pronunciation features from speaker identity, allowing accurate reconstruction of pronunciation (e.g., voice conversion) and coarse-grained pronunciation editing (e.g., foreign accent conversion). In this paper, we demonstrably improve the quality of PPGs to produce a state-of-the-art interpretable PPG representation. We train an off-the-shelf speech synthesizer using our PPG representation and show that high-quality PPGs yield independent control over pitch and pronunciation. We further demonstrate novel uses of PPGs, such as an acoustic pronunciation distance and fine-grained pronunciation control.