SampleRNN

A Tensorflow implementation of SampleRNN: An Unconditional End-to-End Neural Audio Generation Model.

Requirements

• Tensroflow 1.0
• Python 2.7
• Librosa  

Dataset  

We used the pinao music of 74 minutes as the training corpus, and you can use any corpus containing ".wav" files to instead as well.
For Mandarin samples, we used human voice of 6 hours as the training corpus .

Samples

• Mandarin samples
• Pinao samples
• Pinao samples_1
• Pinao samples_2

Pretrained model

• logdir(pinao)
• logdir(mandarin)

FEATURES

• 2-tier SampleRNN
• 3-tier SampleRNN
• Quantization in linear.
• Quantization in mu-law.

Training

python train.py \
	--data_dir=./pinao-corpus \
	--silence_threshold=0.1 \
	--sample_size=102408 \
	--big_frame_size=8 \
	--frame_size=2 \
	--q_levels=256 \
	--rnn_type=GRU \
	--dim=1024 \
	--n_rnn=1 \
	--seq_len=520 \
	--emb_size=256 \
	--batch_size=64 \
	--optimizer=adam \
	--num_gpus=4

or

sh run.sh

Related projects

This work is based on the flowing implementations with some modifications:

• tensorflow-wavenet, a TensorFlow implementation of WaveNet
• sampleRNN_ICLR2017, a Theano implementation of sampleRNN