balacoon/speech_gen_eval_testsets

This repository contains test sets designed to evaluate the performance of speech generation systems. Test sets are compatible with speech_gen_eval and used to collect metrics for TTSLeaderboard.

Repo contains:

• vctk: A 2k subset of the VCTK dataset featuring studio recordings. It is a standard for zero-shot speech generation evaluation. Note that some systems have used this dataset in their training data, which may lead to biased results.
• daps_celeb: A custom dataset combining 1.9k subset of DAPS as source audio and celebrity voices as reference audios. This dataset presents a challenging scenario, simulating a use case where source speech recorded on consumer microphones (e.g., on a phone) is converted to a celebrity voice using a random voice snippet from the internet.