def compute_cmvn(featdir):
#read the spk2utt file
spk2utt = open(featdir + '/spk2utt', 'r')
#create feature reader
reader = ark.ArkReader(featdir + '/feats.scp')
#create writer for cmvn stats
writer = ark.ArkWriter(featdir + '/cmvn.scp')
#loop over speakers
for line in spk2utt:
#cut off end of line character
line = line[0:len(line) - 1]
split = line.split(' ')
#get first speaker utterance
spk_data = reader.read_utt(split[1])
#get the rest of the utterances
for utt_id in split[2:len(split)]:
spk_data = np.append(spk_data, reader.read_utt(utt_id), axis=0)
#compute mean and variance
stats = np.zeros([2, spk_data.shape[1] + 1])
stats[0, 0:spk_data.shape[1]] = np.sum(spk_data, 0)
stats[1, 0:spk_data.shape[1]] = np.sum(np.square(spk_data), 0)
stats[0, spk_data.shape[1]] = spk_data.shape[0]
#write stats to file
writer.write_next_utt(featdir + '/cmvn.ark', split[0], stats)
writer.close()
def decode(self, likelihood, lengths, trans=None):
# Input:
# likelihood: N * L * P numpy array
# N : number of utterance
# L : maximum number of frames of one utterance
# P : dimension of phone posterior (after transform)
# The order should follow lang/phones.txt excepts <eps>,
# #0~#4
# length : 1d numpy array of N lengths of utterances
# trans : Transcription of N utterances, which is a list of N string. If
# None, skip scoring
# Write posterior to feats.scp (In order to call split_data.sh)
writer = ark.ArkWriter(
os.path.join(self.posterior_dir, 'feats.scp'),
os.path.join(self.posterior_dir, 'likelihoods.ark'))
N = likelihood.shape[0]
n_digits = len(str(N))
for idx, (output, l) in enumerate(zip(likelihood, lengths)):
output = output[:l, :]
output = np.where(output == 0, np.finfo(float).eps, output)
output = np.ascontiguousarray(output, dtype=np.float32)
writer.write_next_utt(
str(self._number2str(idx, n_digits)).encode('utf-8'),
np.log(output))
writer.close()
self._gen_utt2spk(N)
if trans:
scoring_cmd = 'false'
self._write_trans(trans)
else:
scoring_cmd = 'true'
os.system(
'%s/scripts/decode.sh --cmd run.pl --skip_scoring %s --nj %s %s %s %s | tee %s/decode.log || exit 1;'
% (os.getcwd(), scoring_cmd, self.nj, self.graph_dir,
self.posterior_dir, self.decode_dir, self.decode_dir))
# Get best WER and print it
wer = os.popen('grep WER %s/wer_* | utils/best_wer.sh' %
self.decode_dir).read()
print(wer)
_, lmwt, penalty = wer[wer.find('wer'):].rstrip().split('_')
output_path = os.path.join(
self.decode_dir,
'scoring_kaldi/penalty_{}/{}.txt'.format(penalty, lmwt))
copy_path = os.path.join(self.decode_dir, 'output.txt')
os.system("cat {} | sort > {}".format(output_path, copy_path))
print(
"The result file of decoding corrsponding to the lowest WER is in: {}\n"
.format(copy_path))
def decode(self, featdir, decodedir):
#create a feature reader
reader = batchdispenser.FeatureReader(featdir + '/feats.scp',
featdir + '/cmvn.scp',
featdir + '/utt2spk',
int(self.conf['context_width']))
#remove ark file if it allready exists
if os.path.isfile(decodedir + '/feats.ark'):
os.remove(decodedir + '/feats.ark')
#open likelihood writer
writer = ark.ArkWriter(decodedir + '/feats.scp')
#create a decoder
decoder = nnetgraph.NnetDecoder(self.DNN, self.input_dim)
#read the prior
prior = np.load(self.conf['savedir'] + '/prior.npy')
#start tensorflow session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(graph=decoder.graph, config=config) as session:
#load the model
decoder.restore(self.conf['savedir'] + '/final')
#feed the utterances one by one to the neural net
while True:
utt_id, utt_mat, looped = reader.getUtt()
if looped:
break
#compute predictions
output = decoder(utt_mat)
#get state likelihoods by dividing by the prior
output = output / prior
#floor the values to avoid problems with log
np.where(output == 0, np.finfo(float).eps, output)
#write the pseudo-likelihoods in kaldi feature format
writer.write_next_utt(decodedir + '/feats.ark', utt_id,
np.log(output))
#close the writer
writer.close()
def prepare_data(datadir, featdir, conf, feat_type, dynamic):
if not os.path.exists(featdir):
os.makedirs(featdir)
#read the segments
if os.path.isfile(datadir + '/segments'):
segments = kaldiInterface.read_segments(datadir + '/segments')
found_segments = True
else:
print(
'WARNING: no segments file found, assuming each wav file is seperate utterance'
)
found_segments = False
#read the wavfiles
wavfiles = kaldiInterface.read_wavfiles(datadir + '/wav.scp')
#create ark writer
writer = ark.ArkWriter(featdir + '/feats.scp')
if os.path.isfile(featdir + '/feats.ark'):
os.remove(featdir + '/feats.ark')
#read all the wav files
RateUtt = {utt: read_wav(wavfiles[utt]) for utt in wavfiles}
#create a featureComputer
comp = feat.FeatureComputer(feat_type, dynamic, conf)
#compute all the features
for utt in wavfiles:
if found_segments:
for seg in segments[utt]:
features = comp(
RateUtt[utt][1][int(seg[1] *
RateUtt[utt][0]):int(seg[2] *
RateUtt[utt][0])],
RateUtt[utt][0])
writer.write_next_utt(featdir + '/feats.ark', seg[0], features)
else:
features = comp(RateUtt[utt][1], RateUtt[utt][0])
writer.write_next_utt(featdir + '/feats.ark', utt, features)
writer.close()
#copy some kaldi files to features dir
copyfile(datadir + '/utt2spk', featdir + '/utt2spk')
copyfile(datadir + '/spk2utt', featdir + '/spk2utt')
copyfile(datadir + '/text', featdir + '/text')
copyfile(datadir + '/wav.scp', featdir + '/wav.scp')
def prepare_data(datadir, featdir, conf, feat_type, dynamic):
'''
compute the features of all segments and save them on disk
Args:
datadir: directory where the kaldi data prep has been done
featdir: directory where the features will be put
conf: feature configuration
featureType: string containing the type of features, optione are:
fbank, mfcc and ssc.
dynamic: the type of dynamic information added, options are:
nodelta, delta and ddelta.
'''
if not os.path.exists(featdir):
os.makedirs(featdir)
#read the segments
if os.path.isfile(datadir + '/segments'):
segments = readfiles.read_segments(datadir + '/segments')
found_segments = True
else:
print( "WARNING: no segments file found, assuming each wav file is seperate utterance")
found_segments = False
#create ark writer
if os.path.isfile(featdir + '/feats.ark'):
os.remove(featdir + '/feats.ark')
writer = ark.ArkWriter(featdir + '/feats.scp', featdir + '/feats.ark')
#read the wavfiles
wavfiles = readfiles.read_wavfiles(datadir + '/wav.scp')
#read all the wav files
rate_utt = {utt: read_wav(wavfiles[utt]) for utt in wavfiles}
#create a featureComputer
comp = feat.FeatureComputer(feat_type, dynamic, conf)
#compute all the features
max_length = 0
for utt in wavfiles:
if found_segments:
for seg in segments[utt]:
features = comp(
rate_utt[utt][1][int(seg[1]*rate_utt[utt][0]):
int(seg[2]*rate_utt[utt][0])],
rate_utt[utt][0])
writer.write_next_utt(featdir + '/feats.ark', seg[0], features)
max_length = max(max_length, features.shape[0])
else:
features = comp(rate_utt[utt][1], rate_utt[utt][0])
writer.write_next_utt(utt, features)
max_length = max(max_length, features.shape[0])
writer.close()
#copy some kaldi files to features dir
copyfile(datadir + '/utt2spk', featdir + '/utt2spk')
copyfile(datadir + '/spk2utt', featdir + '/spk2utt')
copyfile(datadir + '/text', featdir + '/text')
copyfile(datadir + '/wav.scp', featdir + '/wav.scp')
#write the maximum length in a file
with open(featdir + '/maxlength', 'w') as fid:
fid.write(str(max_length))
line = f.readline()
# 将maxlength写入文件
with open(dev_features_dir + "/maxlength", 'w') as f:
f.write("%s"%max_input_length)
print("the utt's maxlength is: " + str(max_input_length))
#create a feature reader
with open(dev_features_dir + '/maxlength', 'r') as fid:
max_length = int(fid.read())
featreader = feature_reader.FeatureReader(dev_features_dir + '/feats.scp', dev_features_dir + '/utt2spk',
context_left, context_right, max_length)
#create an ark writer for the likelihoods
if os.path.isfile(decodedir + '/likelihoods.ark'):
os.remove(decodedir + '/likelihoods.ark')
writer = ark.ArkWriter(decodedir + '/feats.scp', decodedir + '/likelihoods.ark')
#decode with te neural net
nnet.decode(featreader, writer)
print('------- decoding dev sets ----------')
#copy the gmm model and some files to speaker mapping to the decoding dir
os.system('cp %s %s' %(config.get('directories', 'expdir') + '/' + config.get('nnet', 'gmm_name') + '/final.mdl', decodedir))
os.system('cp -r %s %s' %(config.get('directories', 'expdir') + '/' + config.get('nnet', 'gmm_name') + '/graph', decodedir))
os.system('cp %s %s' %(config.get('directories', 'dev_features') + '/utt2spk', decodedir))
os.system('cp %s %s' %(config.get('directories', 'dev_features') + '/text', decodedir))
os.system('cp %s %s' %(config.get('directories', 'dev_features') + '/stm', decodedir))
os.system('cp %s %s' %(config.get('directories', 'dev_features') + '/glm', decodedir))
#change directory to kaldi egs
os.chdir(config.get('directories', 'prjdir'))
