search

chrisdonahue / wavegan

WaveGAN: Learn to synthesize raw audio with generative adversarial networks
MIT License
1.32k stars 283 forks source link

Generate 8 seconds' sample #42

Closed colonder closed 1 year ago

colonder commented 5 years ago

Hi Chris, I'd like to generate 8 seconds' samples with your model. Would you be so kind as to give any hint on how to modify the model?

chrisdonahue commented 5 years ago

Sure, you should be able to modify the WaveGAN generator to support 8 second samples by increasing the starting dim_mul variable:

https://github.com/chrisdonahue/wavegan/blob/master/wavegan.py#L43

Alternatively you can just halve whatever sample rate you're working with using --data_sample_rate to go from 4 to 8 second clips.

jwb95 commented 5 years ago

Hi, I have a modell which I trained with a SLICE_LEN of only 16384 for test purposes. So this modell is capable of generating ~1 second at a sample rate of 16000.

I increased the dim_mul from 16 to 32, 64, 128, 256... without noticing a difference in the length. Halving the sample rate obviously worked to go from ~1 sec up to ~2 sec clips.

Could you give me clearer insight on how to set the dim_mul to achieve longer audio? Since I'm not quite sure what this parameter actually manipulates I'm also wondering how this will affect the quality of the generated audio.

Thanks

chrisdonahue commented 5 years ago

Sorry for the late reply. dim_mul will not change the length of the audio. However, for longer slices (as specified by the data_slice_len parameter), you will likely need to reduce the value of dim_mul or train_batch_size to ensure that the model still fits into memory.

jvel07 commented 4 years ago

Hi, so under this context, what's the difference between data_sample_rate and data_slice_len?

jwb95 commented 4 years ago

data_sample_rate describes the amount of samples used to represent 1 second of audio. cd-standard is 44100 for example. data_slice_len refers to the reception size of the model. e.g. with a data_slice_len of 44100 and s data_sample_rate of 44100 you could let the model learn to generate one second of 44100hz audio.