All projects

Evaluating Accent Conversion Models' (ICPhS 2027, submitted)

Code and stimuli available here: GitHub

This study tested the capability of state-of-the-art voice conversion models to perform accent conversion. These models take in two references: a voice (i.e., timbre) reference and an accent (i.e., style) reference, and output a synthetic voice that sounds like the voice reference speaking in the accent of the accent reference. Methods of testing this type of conversion are coarse and non-standardized. In this study, we tested the efficacy of these models using a careful perception study. We used OpenVoice and Seed-VC. What we found was that Seed-VC performs better at preserving the voice of the input but does not definitively transfer accent. OpenVoice, on the other hand, successfully transfers accent but does not preserve the voice of the input.

Examples: Listen for yourself

Both models are trained to produce an output that preserves the voice of the voice reference while transferring the accent of the accent reference.

OpenVoice

Voice reference

Accent reference

Output

Seed-VC

Voice reference

Accent reference

Output

Bar charts comparing how often OpenVoice and Seed-VC outputs were judged to have the same accent as the voice or accent reference
Results of the accent similarity experiment. Participants were presented with two short audio files and asked, “Do these two have the same accent?” Trial type is shown on the y-axis. Output refers to model output, voice refers to the corresponding voice reference, and accent refers to the corresponding accent reference. OpenVoice preserves accent. Seed-VC seems to produce something that’s ambiguous in terms of accent.
Bar charts comparing how often OpenVoice and Seed-VC outputs were judged to have the same speaker identity as the voice or accent reference
Results of the speaker identity experiment. Participants were presented with two short audio files and asked, “Do these two have the same speaker?” Trial type is shown on the y-axis. Output refers to model output, voice refers to the corresponding voice reference, and accent refers to the corresponding accent reference. Seed-VC preserves voice.