A COMPARISON OF DISCRETE AND SOFT SPEECH UNITS FOR IMPROVED VOICE CONVERSION

Authors: Benjamin van Niekerk, Marc-André Carbonneau, Julian Zaïdi, Matthew Baas, Hugo Seuté, Herman Kamper

Abstract: The goal of voice conversion is to transform source speech into a target voice, keeping the content unchanged. In this paper, we focus on self-supervised representation learning for voice conversion. Specifically, we compare discrete and soft speech units as input features. We find that discrete representations effectively remove speaker information but discard some linguistic content - leading to mispronunciations. As a solution, we propose soft speech units. To learn soft units, we predict a distribution over discrete speech units. By modeling uncertainty, soft units capture more content information, improving the intelligibility and naturalness of converted speech.

If you are having trouble listening to the audio, try refreshing the page.

Intra-lingual - English

In this section, we present some speech samples used in the intra-lingual subjective evaluation.
We focus on any-to-one conversion using LJSpeech as the target and LibriSpeech dev-clean as the source speech.
We compare discrete and soft speech units as well as two baselines.

Target speaker:

Source	HuBERT-Soft	HuBERT-Discrete	CPC-Soft	CPC-Discrete	Cascaded ASR-TTS	AutoVC

Cross-lingual - French

In this section, we present some cross-lingual speech samples for French.
We use LJSpeech as the target and CSS10 as source speech.
We compare HuBERT-soft and HuBERT-discrete.

Target speaker:

Source	HuBERT-Soft	HuBERT-Discrete

Cross-lingual - Afrikaans

In this section, we present some cross-lingual speech samples for Afrikaans (one of South Africa's official languages).
We use LJSpeech as the target and a South African languages corpus as source speech.
We compare HuBERT-soft and HuBERT-discrete.

Target speaker:

Source	HuBERT-Soft	HuBERT-Discrete