labels = []
label_dir = '/home/pony/github/ASR_libri/libri/cha-level/mfcc_and_label/label/'
mfcc_dir = '/home/pony/github/ASR_libri/libri/cha-level/mfcc_and_label/mfcc/'
if True:
for subdir, dirs, files in os.walk(rootdir):
if True:
for f in files:
fullFilename = os.path.join(subdir, f)
filenameNoSuffix = os.path.splitext(fullFilename)[0]
if f.endswith('.wav'):
print fullFilename
(rate, sig) = wav.read(fullFilename)
mfcc = calcMFCC_delta_delta(sig,
rate,
win_length=0.020,
win_step=0.020)
# transpose mfcc to array of (39,time_length)
mfcc = np.transpose(mfcc)
print mfcc.shape
# save mfcc to file
m_f = mfcc_dir + filenameNoSuffix.split('/')[-1] + '.npy'
np.save(m_f, mfcc)
labelFilename = filenameNoSuffix + '.label'
with open(labelFilename, 'r') as f:
characters = f.readline().strip()
print characters
targets = []
for c in characters:
if c == ' ':
def wav_to_mfcc(wav_file_path):
rate, sig = wav.read(wav_file_path)
mfcc = calcMFCC_delta_delta(sig, rate, win_length=0.020, win_step=0.010)
mfcc = np.transpose(mfcc)
return mfcc
output_logits = model(feat)
output_probs = nn.functional.softmax(output_logits, dim=1)
output_label = int(np.argmax(output_probs.squeeze().cpu().numpy()))
output_probs = output_probs.squeeze().tolist()
print("output_logits:{}".format(output_logits[0][output_label]))
end_time=time.time()
print("Model predicting time {}".format(end_time - start_time))
#print(output_logits)
#print(output_probs)
#print(output_label)
output_person = label2person[output_label]
return output_probs, output_label, output_person
if __name__=="__main__":
import scipy.io.wavfile as wav
from calcmfcc import calcMFCC_delta_delta
path = "./storage/Qiaomu.wav"
(rate,sig) = wav.read(path)
mfcc_feat = calcMFCC_delta_delta(sig,rate)
model = RNN_model(14)
model.load_state_dict(torch.load("models/best_model.pt"))
with open("models/label_person_mapping.pickle", "rb") as file:
label_person_mapping = pickle.load(file)
label2person = label_person_mapping["label2person"]
print(label2person)
output_prob, output_person = classify_one_cycle(model, mfcc_feat, label2person)
print("output_person:{}".format(output_person))
def POST(self):
web.header('Access-Control-Allow-Origin', '*')
web.header('Access-Control-Allow-Credentials', 'true')
web.header('Access-Control-Allow-Methods', 'POST, GET, OPTIONS')
data = web.input()
start_time = []
end_time = []
sil_start_time = []
sil_end_time = []
if data.type == 'audio/mpeg':
wf = open(dir_ + 'storage/' + str(data.url)[-6:] + '.mp3', 'wb')
wf.write(data.file)
wf.close()
elif data.type == 'audio/wav':
#read sample and create a copy
sample_name = data.filename
sf = open(dir_ + 'storage/' + sample_name, 'rb')
wav_bytes = sf.read()
file_name = 'curr.wav'
wf = open(dir_ + 'storage/' + file_name, 'wb')
wf.write(wav_bytes)
wf.close()
print("audio received ...")
# make single channel [uncomment if passing dual channel audio]
#silence_classf.make_one_channel(dir_+'storage',file_name)
# call voice separation required coz due to ambient noise the silence detection doesn't work properly
#new_fn_bg,new_fn_fg=silence_classf.vocal_separation(dir_+'storage',file_name)
#sil_file_name=new_fn_fg
# call umm segmentation
(rate, sig) = wav.read(dir_ + 'storage/' + file_name)
start_time = time.time()
mfcc_feat = calcMFCC_delta_delta(sig, rate)
mid_time = time.time()
print("Signal processing time:{} s".format(mid_time - start_time))
model = RNN_model(15)
if torch.cuda.is_available():
device = torch.device('cuda:0')
else:
device = torch.device('cpu')
model.load_state_dict(
torch.load("models/best_model.pt", map_location=device))
with open("models/label_person_mapping_new.pickle", "rb") as file:
label_person_mapping = pickle.load(file)
label2person = label_person_mapping["label2person"]
output_prob, output_label, output_person = classify_one_cycle(
model, mfcc_feat, label2person)
end_time = time.time()
target_prob = output_prob[output_label]
if "_" in output_person:
output_person = output_person.split(
"_")[0] + " " + output_person.split("_")[1]
print("output_person:{}".format(output_person))
print("output_probabilities:{}".format(target_prob))
print("Model predicting time:{} s".format(end_time - mid_time))
return json.dumps({
'msg': "Your file is uploaded!",
'output_person': output_person,
"output_probability": round(target_prob, 4),
"total_time": round(end_time - start_time, 3),
})