search

agostini01 / Real-Time-Voice-Cloning

Clone a voice in 5 seconds to generate arbitrary speech in real-time
0 stars 4 forks source link

Voice Encoder #5

Open agostini01 opened 4 years ago

agostini01 commented 4 years ago
malithj commented 4 years ago

Finished quantization.

Image of Quantized Encoder - Speedup

Image of Quantized Encoder - Storage

agostini01 commented 4 years ago

Wow!

@malithj , This is rather impressive! Really cool.

Just to confirm as you have much more experience with quantization than I do. Are the following steps correct?

  1. Isolate the model and parameters
  2. Load the model weights
  3. Use torch.quantization.quantize_* functions to perform the quantization - passing in the correct datatype.

What datatype did you use?

Really awesome work! Can it still do the same kind of encoding?

malithj commented 4 years ago

Thanks.

Yep. I used torch.qint8 so that there would be 256 (2^8) quantization levels. The sound quality doesn't seem to degrade that much.

Here's the script I used. https://github.com/malithj/Deep-learning/blob/master/FinalProject/psq.py

But pytorch only supports quantization on CPUs. So, testing will have to be done using CPUs.

agostini01 commented 4 years ago

Understood! Sounds reasonable!