def __init__(self, samplingFrequency=8000, framePeriod=25e-3, hopPeriod=10e-3, trainDir="./train_audio/",
thresh=1.2):
self.samplingFrequency = samplingFrequency
self.framePeriod = framePeriod
self.hopPeriod = hopPeriod
self.trainDir = trainDir
self.hopLength = int(samplingFrequency * hopPeriod)
self.frameLength = int(samplingFrequency * framePeriod)
self.referenceMFCC = []
for file_name in os.listdir(trainDir):
if file_name.endswith(".wav"):
(fs, data) = wv.read(trainDir + file_name)
num_frames = int(data.shape[0] / self.hopLength) - int(np.ceil(self.frameLength / self.hopLength))
MFCC_calculator = mfcc.MFCC()
MFCC_MATRIX = np.empty([39, abs(num_frames)])
for k in range(num_frames):
MFCC_MATRIX[:, k] = MFCC_calculator.compute_mfcc(
data[k * self.hopLength: k * self.hopLength + self.frameLength])
self.referenceMFCC.append(MFCC_MATRIX)
DTW_calculator = dtw.DTW()
distance_list = [DTW_calculator.compute_distance(np.transpose(matrix), np.transpose(matrix2)) for matrix in
self.referenceMFCC for matrix2 in self.referenceMFCC]
self.thresh = np.mean(np.array((distance_list))) * thresh
def distance(self, fileName):
"""
This function is used for calculating the DTW distance between the test utterance and each of the 10 training utterances.
:param fileName: Name of test utterance .wav file
:type fileName: str
:returns: List of DTW distances of test utterance with each training utterance
:rtype: list
"""
if isinstance(fileName, (bytes, bytearray)):
data = np.fromstring(fileName)
else:
(fs, data) = wv.read(fileName)
DTW_calculator = dtw.DTW()
num_frames = int(data.shape[0] / self.hopLength) - int(np.ceil(self.frameLength / self.hopLength))
if num_frames <= 0:
return 10000
MFCC_calculator = mfcc.MFCC()
MFCC_MATRIX = abs(np.empty([39, num_frames]))
for k in range(num_frames):
MFCC_MATRIX[:, k] = MFCC_calculator.compute_mfcc(
data[k * self.hopLength: k * self.hopLength + self.frameLength])
distance_list = [DTW_calculator.compute_distance(np.transpose(matrix), np.transpose(MFCC_MATRIX)) for matrix in
self.referenceMFCC]
return distance_list
def main(args):
words = []
wordstoindex = {}
mfcc_convert = mfcc.MFCC()
mfcc_savepath_prefix = os.path.splitext(args.mfccs)[0]
num_frames = 0
h5file = h5py.File(args.mfccs, "w")
for line_id, line in enumerate(open(args.wordlist, "r").readlines()):
word, path, start, end = line.strip().split()
if not wordstoindex.has_key(word):
wordstoindex[word] = len(words)
words.append(word)
class_id = wordstoindex[word]
start, end = int(start), int(end)
command = ("sox -t sph {0} -r 16000 -t wav {1}"
).format(path, "tmp.wav")
subprocess.call(command, shell=True)
sample_rate, samples = wavfile.read("tmp.wav")
samples = samples.astype(float)
full_mfccs = mfcc_convert.sig2s2mfc(samples)
cutoff = numpy.sort(full_mfccs[:, 0])
speech_areas = full_mfccs[full_mfccs[:, 0] > cutoff]
cepstral_mean = speech_areas.mean(axis=0)
cepstral_variance = speech_areas.std(axis=0)
word_samples = samples[start - 480: end + 480]
word_mfccs = ((mfcc_convert.sig2s2mfc(word_samples) - cepstral_mean)
/ cepstral_variance)
n_frames, n_dims = word_mfccs.shape
full_mfccs = numpy.empty((n_frames, n_dims * 3), dtype=numpy.float32)
full_mfccs[:,:n_dims] = word_mfccs
mfcc.deltas(word_mfccs, output_frames=full_mfccs[:,n_dims:2*n_dims])
mfcc.deltas(full_mfccs[:,n_dims:2*n_dims],
output_frames=full_mfccs[:,2*n_dims:])
if line_id == 0:
mfcc_dset = h5file.create_dataset("mfccs", (n_frames, 39), maxshape=(None, 39), dtype=numpy.float32)
label_dset = h5file.create_dataset("labels", (100, 3), maxshape=(None, 3), dtype=numpy.int32)
data_idx = 0
cur_idx = 0
if cur_idx + n_frames >= len(mfcc_dset):
h5file.flush()
mfcc_dset.resize((2*len(mfcc_dset), mfcc_dset.shape[1]))
print "mfcc doubling", cur_idx
if data_idx == len(label_dset):
h5file.flush()
label_dset.resize((2*label_dset.shape[0], label_dset.shape[1]))
try:
mfcc_dset[cur_idx:cur_idx+n_frames] = full_mfccs
except:
import pdb; pdb.set_trace()
label_dset[data_idx] = class_id, cur_idx, n_frames
cur_idx += n_frames
data_idx += 1
mfcc_dset.resize((cur_idx, mfcc_dset.shape[1]))
label_dset.resize((data_idx, label_dset.shape[1]))
h5file.flush()
words.sort()
open(args.wordkey, "w").write("\n".join(
"%s %d" % (w, wordstoindex[w]) for w in words))
def process_one_file(wav_base_name):
wavefilename = wavdir + "/" + wav_base_name + ".wav"
print wavefilename
fh = wave.open(wavefilename, "r")
sampwidth = fh.getsampwidth()
print fh.getparams()
nsamples = fh.getnframes()
bytes = fh.readframes(nsamples)
code = "h"
samples = array.array(code, bytes)
samples = np.array(samples)
fh.close()
print len(samples)
print samples[12000:12020]
mfcc_processor = mfcc.MFCC()
feats = mfcc_processor.sig2s2mfc(samples)
print len(feats)
print feats[10:12]
featfilename = featdir + "/" + wav_base_name + ".feat"
feats.dump(featfilename)
checkfeats = np.load(featfilename)
assert (feats == checkfeats).all()