How to evaluate the quality of synthesized speech？ about deepvoice3_pytorch HOT 4 CLOSED

I'm not so familiar with evaluation metrics for TTS. You can check out scientific papers here: https://research.google.com/pubs/HeigaZen.html.

from deepvoice3_pytorch.

@zuoxiang95
This might be helpful:
http://tts.speech.cs.cmu.edu/courses/11492/slides/tts_eval.pdf

You could use MCD(Mean Cepstral distortion) as an objective metric..(its a weighted Euclidean distance between true and synthesized acoustic vector).. assumes you have ground truth data.
mcd implementation between two Mceps:
https://github.com/festvox/festvox/blob/master/src/clustergen/track_diff_mcd

MCD implementation between two wavfiles using DTW:
https://github.com/festvox/festvox/blob/master/src/clustergen/get_cd_dtw

Other Implementation:
https://github.com/MattShannon/mcd

Alternatively, you could use SUS (Semantically Unpredictable sentences) for Understandability judgement asking users to transcribe SUS sentences and measure word error rate.

from deepvoice3_pytorch.

You could also use word error rate

from deepvoice3_pytorch.

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

from deepvoice3_pytorch.