Synthetic speech

#6
by cdminix - opened

Quite a few of the English samples are synthetic speech, does anyone know a good way to filter those out?

Amphion org

Hi thanks for the question, maybe this scripts help https://github.com/SWivid/F5-TTS/blob/main/scripts/prepare_emilia.py.

Were the specific synthetic file IDs in the prepare script excluded from the dataset?
In my sample of 200 clips almost 20% were likely synthetic...

Amphion org

This script excludes certain audio files that appear to be synthesized or heavily code-switched, which could still be relevant for your use case. You may refer to our source audio list at this link (https://huggingface.co/datasets/amphion/Emilia), where tags indicating whether an audio file is synthesized are available. For more robust filtering, it is recommended to train a state-of-the-art (SOTA) anti-spoofing model, such as the one available here (https://github.com/TakHemlata/SSL_Anti-spoofing), which provides a pre-trained checkpoint (ckpt). You can then use this model to infer over the entire dataset and filter out audio files that are confidently identified as synthesized.

Thank you very much, that resolves it for me :)

cdminix changed discussion status to closed

Sign up or log in to comment