Not able to get good audio quality

[alexdemartos]

Hi,

I am training a WaveNet model from Rayhane-mamah Tacotron-2 GTA mels (I've modified mel2samp_onehot.py accordingly) and I am not able to get good audio quality.

This is the audio I get from a training sample after 250k iter (batch size 7):
original.wav
250k_iter.wav

I know I might want to train longer, but I've already trained a 1.3M iter WaveNet model and I get the same "low quality" audio.

Here are my parameters:

"data_config": {
        "segment_length": 22050,
        "mu_quantization": 256,
        "filter_length": 2048,
        "hop_length": 275,
        "win_length": 1100,
        "sampling_rate": 22050,
        "mel_fmin": 50,
        "mel_fmax": 7600
    },
    "wavenet_config": {
        "n_in_channels": 256,
        "n_layers": 16,
        "max_dilation": 256,
        "n_residual_channels": 64,
        "n_skip_channels": 256,
        "n_out_channels": 256,
        "n_cond_channels": 80,
        "upsamp_window": 1100,
        "upsamp_stride": 275
    }

Any clues what might be wrong? Maybe some pre/post processing missing? Maybe this is due to mulaw-quantize with 256 values?

Any help is really appreaciated. Thank you very much.

[WhiteFu]

hello!
Have you solved the problem?