ppg training about ppg-vc

[liusongxiang]

Please refer to Espnet for the trainig process.

[Lukelluke]

Please refer to Espnet for the trainig process.

Hi, Dr.liu,

Your job is so gorgeous, use the Encoder part of ASR task instead of traditional Kaldi way, this inspires us much!

We explored ESPNet, but still have some questions wanna get ur help:

Q1: You provided the “/conformer_ppg_model/*” files, and if we wanna make these runnable as the "espnet/egs/librispeech/asr1/" example in ESPNet, how should we make preparation steps? For instance, 1、the data preparation, 2、the files organization, and 3、how to prepare the corresponding "run.sh" script?

Q2: As ur description in paper, the bottle neck features of Encoder output in ASR task ,are extracted as "speaker independent information". Do this kind of features can be equal to traditional "ppg" features? Further more, can we researchers work in Voice Conversion field, take this way to extract "ppg" features, instead of traditional Kaldi way?

Best wishes!
Luke

[liusongxiang]

Thank you for the questions.
For Q1:
I adapted espnet a lot; it seems that espnet asr models always downsample the encoder input along the temporal axis more than 4x and do not support phoneme as output symbols. Source codes should be modified correspondingly for VC applications. But the basic steps for the training process is very similar to those presented in espnet asr recipes, including the data preparation, files organization. The run.sh should be modified a little bit, e.g., the language model can be skipped. Sufficient familiarity of espnet source code should be necessary if you want to train a content encoder using your own data.

For Q2:
Please refer to this paper for your questions: TTS Skins: Speaker Conversion via ASR.
Good VC performance validate the speaker independence property of the bottle neck feature obtained in this way. The paper listed above says that BNF is better than PPG features, but this could really be a model selection thing.

Hope this can help.
Songxiang Liu