Script for preprocessing, take spk p251 vs p299 for example.

preprocess.py

import os, sys
import numpy as np
import h5py
from collections import defaultdict

from hyperparams import Hyperparams as hp
from utils import *

h5py_path = sys.argv[1]
dataset_path = sys.argv[2]

p251_mcs = []
p251_logf0s = []
p299_mcs = []
p299_logf0s = []

# extracting
for dirPath, dirNames, fileNames in os.walk(dataset_path):
    for f in fileNames:
        ids = f.replace('.wav', '')[1:]
        speaker_id, uttr_id = ids.split('_')
        fname = os.path.join(dirPath, f)
        print('processing %s' % fname)
        y, sr = librosa.load(fname, sr=hp.sr, dtype=np.float64)
        mc, f0, ap = get_MCEPs(y)
        logf0 = np.log(f0 + 1.0)
        if speaker_id == '251':
            p251_mcs.extend(mc)
            p251_logf0s.extend(logf0)
        else:
            p299_mcs.extend(mc)
            p299_logf0s.extend(logf0)

p251_mc_mean = np.mean(np.array(p251_mcs), axis=0)
p251_mc_std = np.std(np.array(p251_mcs), axis=0)
p251_logf0_mean = np.mean(np.array(p251_logf0s))
p251_logf0_std = np.std(np.array(p251_logf0s))

p299_mc_mean = np.mean(np.array(p299_mcs), axis=0)
p299_mc_std = np.std(np.array(p299_mcs), axis=0)
p299_logf0_mean = np.mean(np.array(p299_logf0s))
p299_logf0_std = np.std(np.array(p299_logf0s))


# start to make h5py
with h5py.File(h5py_path, 'w') as f_h5:
    for dirPath, dirNames, fileNames in os.walk(dataset_path):
        for f in fileNames:
            ids = f.replace('.wav', '')[1:]
            speaker_id, uttr_id = ids.split('_')
            
            if speaker_id == '251':
                mc_mean = p251_mc_mean
                mc_std = p251_mc_std
                logf0_mean = p251_logf0_mean
                logf0_std = p251_logf0_std
            else:
                mc_mean = p299_mc_mean
                mc_std = p299_mc_std
                logf0_mean = p299_logf0_mean
                logf0_std = p299_logf0_std
            
            fname = os.path.join(dirPath, f)
            y, sr = librosa.load(fname, sr=hp.sr, dtype=np.float64)
            mc, f0, ap = get_MCEPs(y)
            normed_mc = np.expand_dims((mc - mc_mean) / mc_std, axis=0)
            normed_logf0 = np.expand_dims(((np.log(f0 + 1.0)) - logf0_mean) / logf0_std, axis=0)
            ap = np.expand_dims(ap, axis=0)
            
            # features
            f_h5.create_dataset('{}/{}/normed_mc'.format(speaker_id, uttr_id), data=np.array(normed_mc),
                dtype=np.float64)
            f_h5.create_dataset('{}/{}/normed_logf0'.format(speaker_id, uttr_id), data=np.array(normed_logf0),
                dtype=np.float64)
            f_h5.create_dataset('{}/{}/ap'.format(speaker_id, uttr_id), data=np.array(ap),
                dtype=np.float64)
                   
            mc_mean = np.expand_dims(mc_mean, axis=0)
            mc_std = np.expand_dims(mc_std, axis=0)
            logf0_mean = np.expand_dims(logf0_mean, axis=0)
            logf0_std = np.expand_dims(logf0_std, axis=0)
            
            # statistics
            f_h5.create_dataset('{}/{}/mc_mean'.format(speaker_id, uttr_id), data=np.array(mc_mean),
                dtype=np.float64)
            f_h5.create_dataset('{}/{}/mc_std'.format(speaker_id, uttr_id), data=np.array(mc_std),
                dtype=np.float64)
            f_h5.create_dataset('{}/{}/logf0_mean'.format(speaker_id, uttr_id), data=np.array(logf0_mean),
                dtype=np.float64)
            f_h5.create_dataset('{}/{}/logf0_std'.format(speaker_id, uttr_id), data=np.array(logf0_std),
                dtype=np.float64)