Skip to content

Latest commit

 

History

History
58 lines (48 loc) · 1.83 KB

dataset.md

File metadata and controls

58 lines (48 loc) · 1.83 KB

Dataset

This document describe how to prepare the dataset.

A. Preparation

Due to the copyright issue, we cannot distrubute the audio files and the mel-spectrograms of MPop600[1] dataset. However, readers can following the steps to build one from scratch.

1. File Strutcure

Organized the dataset as the following (or similar) structure.

data          
├─ f1                 
├─ m1     
│  ├─ test
│  ├─ train-3min
│  ├─ train-full
│  ├─ train-val
│  │  ├─  audio
│  │  │  ├─  xxx.wav
│  │  ├─  mel
│  │  │  ├─  xxx.npy

In this example, we have one female (f1) and a male (m1) singer. For each singer, we have 2 training sets for different scenatios: train-full (6 hours), and train-3min (3 minutes only). The two scenatios use the same validation (train-val) and testing set (test).

The specification of the audio files:

  • format: .WAV
  • Sampling rate: 24kHz
  • Bit depth: 16bit

2. Preprocessing

Run preprocess.py to generate mel-spetrogram files, which is stored in .npy format. Remember to modify the configuration in preprocess.py at first.

'''in preprocess.py'''

# ...
# ==================================================== #
# configuration
# ==================================================== #
path_rootdir = './data'
device = 'cuda'

sampling_rate  = 24000
hop_length     = 240
win_length     = 1024
n_mel_channels = 80

src_ext = 'wav'
dst_ext = 'npy'
# ...

3. Configuration

Before training, remember to change the dataset path in the configuration file, which can be founder under configs folder and stored in .yaml format.

B. Reference

[1] (APSIPA ASC'20) MPop600: A Mandarin Popular Song Database with Aligned Audio, Lyrics, and Musical Scores for Singing Voice Synthesis