Giters

bshall / UniversalVocoding

A PyTorch implementation of "Robust Universal Neural Vocoding"

Home Page:https://bshall.github.io/UniversalVocoding/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Inference speed comparison

sujeendran opened this issue · comments

commented

Hi! Could you share some details about the inference speed compared to Griffin-Lim/WaveNet/WaveRNN?

commented

Hi @sujeendran,

Sorry about the delay. With a GeForce GTX 1080 Ti I'm getting around 3700 samples a second (so almost 0.25x real-time with 16kHz audio) and on an Intel Core i7-8700K CPU @ 3.70GHz I'm getting about 1700 samples a second.

So definitely faster than vanilla WaveNet but not quite real-time. However, since there is only a single forward GRU on the autoregressive path I'm sure you could get better than real-time with some engineering e.g. sparsifying the GRU, designing a custom cuda kernel, etc.