Hi, thanks for your great work. After reading your paper, I have a question here.<

Hi, <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url=

Hi <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="

some questions about voicefixer HOT 2 CLOSED

LqNoob commented on May 20, 2024

some questions

from voicefixer.

Comments (2)

LqNoob commented on May 20, 2024 1

Hi, @haoheliu, thank you for your quick and kindly reply.
Because the two-stage model algorithm may bring the burden of model size, so I will consider the model's inference speed. Now your answer helped me a lot.
Thank you.

from voicefixer.

haoheliu commented on May 20, 2024

Hi @LqNoob, thanks for the questions.

Because I found vocoder is very useful especially in speech super-resolution task. In order to utilize its power, I use the two-stage algorithm. Also processing on mel spectrogram is easier than linear spec.
The training of the analysis module (VF-UNet) takes about two days on 4-V100-32G GPUs. Actually, we prioritize performance over model size and inference speed so didn't pay too much attention to computational overhead. I suppose it can be optimized easily with lighter architecture.

from voicefixer.

Recommend Projects