Evaluating how neural vocoders generalize to unseen pitch values rather than only testing within their training conditions.
Time seriesSpeech synthesis
Conceptual illustration. Training data are split to control seen and unseen pitch values, and different vocoder families are evaluated for extrapolation beyond their training conditions. Open full size ↗
This research introduces a methodology for benchmarking the extrapolation capabilities of neural vocoders by controlling which pitch values are seen during training. It compares autoregressive, source-filter, flow, and GAN approaches.
Conducted during my research internship at GIPSA-LAB in Grenoble. See the Interspeech publication
for the method and results.
