用CatBoost做音素识别

1，下载TIMIT语料库，解压，全部目录，文件变小写，按照NLTK要求，修改所有根目录下的.txt配置文件

2，安装必要python库：nltk， librosa，numpy，catboost

3,  实现训练过程，原理：TIMIT中wav格式并非标准RIFF格式，好在格式简单，每个文件头占1024字节，文本格式，剩下的是pcm raw data，录制格式是单声道，16bit 16KHz，按照数据格式直接填入numpy.array里面，然后交由librosa，提取MFCC以及MFCC delta特征，做为训练输入数据，同时分析同名的phn文件，生成对应的训练标签。

import nltk
from nltk.corpus import timit
import numpy as np 
import librosa
import os
from catboost import CatBoostClassifier


def exactFeature(phones, item):
    wavfile = "{}/{}.wav".format(timit._root, item)
    """
    se_version -s3 1.0
    utterance_id -s10 aks0_sx133
    channel_count -i 1
    sample_count -i 53044
    sample_rate -i 16000
    sample_min -i -3305
    sample_max -i 3700
    sample_n_bytes -i 2
    sample_byte_format -s2 01
    sample_sig_bits -i 16
    end_head
    """

    f = open(wavfile, "rb")
    f.seek(1024, os.SEEK_SET)
    d = np.fromfile(f, dtype=np.int16).astype(np.float32)
    sr = 16000
    mfccs = librosa.feature.mfcc(d, sr, n_mfcc=13, hop_length=int(0.010*sr), n_fft=int(0.025*sr)) 
    mfccs[0] = librosa.feature.rmse(d, hop_length=int(0.010*sr), n_fft=int(0.025*sr)) 
    deltas = librosa.feature.delta(mfccs) 
    mfccs_plus_deltas = np.vstack([mfccs, deltas]) 
    dataset=np.flipud(np.rot90(mfccs_plus_deltas))
    curpos=0
    train_labels=[]
    for v in timit.phone_times(item):
        while v[2]>curpos:
            train_labels.append(phones.index(v[0]))
            curpos+=0.010*sr
    train_labels.insert(0,train_labels[0])
    for ind in xrange(len(train_labels),dataset.shape[0]):
        train_labels.append(phones.index(v[0]))
    train_labels = train_labels[:dataset.shape[0]]
    return dataset,train_labels

phones=list(set(timit.phones()))

alldataset=np.array([])
alltrain_lab=[] 
for item in timit.utteranceids():
    dataset,train_labels=exactFeature(phones,item)
    if alldataset.shape[0]==0:
        alldataset = dataset
    else:
        alldataset=np.concatenate((alldataset, dataset))
    alltrain_lab.extend(train_labels)
    if alldataset.shape[0] != len(alltrain_lab):
        print "feature no match!!!", dataset.shape, len(train_labels), alldataset.shape, len(alltrain_lab)
        break
model = CatBoostClassifier(learning_rate=0.1, depth=12, loss_function='MultiClass')
fit_model = model.fit(alldataset, alltrain_lab)
print fit_model.get_params()

4， 样本数量：1941232，在单台centos虚拟机上，纯cpu训练过程不超过1小时，的确非常高效

训练过程如下：