speakers = ['fuli', 'lzh', 'cl', 'lsq', 'lxx', 'wry', 'lms', 'zlb', 'zq']
enroll_wav_path = 'media'
dict_spkid_embeddings = {}
enroll_nums = 5
total_wavs = glob.glob(os.path.join(data, '*.wav'))
print('total wavs: ', len(total_wavs))
for speaker in speakers:
#print(speaker)
speaker_wavs = glob.glob(os.path.join(enroll_wav_path,
speaker + '_*.wav'))[:5]
#print(speaker, speaker_wavs)
speaker_embeddings = []
for wav in speaker_wavs:
mfcc = sample_from_mfcc(read_mfcc(wav, SAMPLE_RATE), NUM_FRAMES)
predict_feat = model.m.predict(np.expand_dims(mfcc, axis=0))
speaker_embeddings.append(predict_feat)
num_utterances = len(speaker_wavs)
enroll_centroid_embeddings = get_centroid(speaker_embeddings,
num_utterances)
dict_spkid_embeddings[speaker] = enroll_centroid_embeddings
a = 0
total = 0
score = 0
name = 'Who?'
test_path = 'media'
for speaker in speakers:
test_wavs = glob.glob(os.path.join(test_path, speaker + '_*.wav'))[6:]
def record(self):
####计算enroll centeriod embedding####
dict_spkid_embeddings = {}
all_wavs = glob.glob(os.path.join(enroll_wav_path, '*.wav'))
model = DeepSpeakerModel()
#model.m.load_weights('weights/ResCNN_triplet_training_checkpoint_265.h5', by_name=True)
model.m.load_weights(
'weights/ResCNN_softmax_pre_training_checkpoint_102.h5',
by_name=True)
## speakers###
speakers = ['lms', 'zq', 'wry', 'lzh']
# enroll_centroid_embeddings_npy = np.zeros((3, 256),dtype=float)
# enroll_dict= {}
i = 0
for speaker in speakers:
#print(speaker)
speaker_wavs = glob.glob(
os.path.join(enroll_wav_path, speaker + '_*.wav'))[:6]
#print(speaker_wavs)
speaker_embeddings = []
for wav in speaker_wavs:
mfcc_feat = sample_from_mfcc(read_mfcc(wav, SAMPLE_RATE),
NUM_FRAMES)
mfcc_feat = model.m.predict(np.expand_dims(mfcc_feat, axis=0))
speaker_embeddings.append(mfcc_feat)
num_utterances = len(speaker_wavs)
enroll_centroid_embeddings = get_centroid(speaker_embeddings,
num_utterances)
dict_spkid_embeddings[speaker] = enroll_centroid_embeddings
i += 1
#np.save('enroll_3.npy', enroll_dict)
score_speakers = {}
thres = 0.10
self.pause_flag = False
#self.record_pushButton.setText('停止')
# 创建PyAudio对象
self.pa = PyAudio()
# 打开声卡，设置 采样深度为16位、声道数为2、采样率为16、模式为输入、采样点缓存数量为2048
stream = self.pa.open(format=paInt16,
channels=1,
rate=16000,
input=True,
frames_per_buffer=3200)
# 新建一个列表，用来存储采样到的数据
record_buf = []
while True:
if self.pause_flag is True:
break
audio_data = stream.read(3200) # 读出声卡缓冲区的音频数据
record_buf.append(audio_data) # 将读出的音频数据追加到record_buf列表
my_path = 'test/' + strftime("%Y%m%d%H%M%S", localtime(
time())) + '.wav'
wf = wave.open(my_path, 'wb') # 创建一个音频文件
wf.setnchannels(1) # 设置声道数为2
wf.setsampwidth(2) # 设置采样深度为
wf.setframerate(16000) # 设置采样率为16000
# 将数据写入创建的音频文件
wf.writeframes("".encode().join(record_buf))
# 写完后将文件关闭
wf.close()
# 停止声卡
stream.stop_stream()
# 关闭声卡
stream.close()
# 终止pyaudio
self.pa.terminate()
self.pa = None
#self.record_pushButton.setText('录制')
self.record_pushButton.setEnabled(True)
test_wav = my_path
mfcc_feat = sample_from_mfcc(read_mfcc(test_wav, SAMPLE_RATE),
NUM_FRAMES)
output_feat = model.m.predict(np.expand_dims(mfcc_feat, axis=0))
score = 0
name = 'Who?'
for speaker_name in dict_spkid_embeddings.keys():
score_speaker = batch_cosine_similarity(
dict_spkid_embeddings[speaker_name], output_feat)
print('speaker: ', speaker_name, 'score: ', score_speaker)
if score_speaker > score:
score = score_speaker
name = speaker_name
print('speaker: ', name)
print('score: ', score)
self.speaker_label.setText('Speaker: %s' % name)
self.score_label.setText('Score: %.4f' % score)
random.seed(123)
# Define the model here.
model = DeepSpeakerModel()
# Load the checkpoint.
model.m.load_weights('weights/ResCNN_triplet_training_checkpoint_265.h5',
by_name=True)
data = 'data_eng'
#labels = ['cl', 'fuli', 'gongwenhua', 'liuyuguang', 'lms', 'lsq', 'lxx', 'lzh', 'shanke', 'wry', 'zhangshuai163', 'zhuting', 'zlb', 'zq']
labels = ['fuli', 'lzh', 'cl', 'lsq', 'lxx', 'wry', 'lms', 'zlb', 'zq']
for speaker in labels:
wavs = glob.glob(os.path.join(data, speaker, '*.wav'))
print('{}: {} wavs'.format(speaker, len(wavs)))
if (__name__ == '__main__'):
print('Load model successfully.')
inputs_wavs = glob.glob(os.path.join(data, '*', '*.wav'))
print(len(inputs_wavs))
for wav in inputs_wavs:
#print(wav)
mfcc = sample_from_mfcc(read_mfcc(wav, SAMPLE_RATE), NUM_FRAMES)
#print(mfcc.shape)
predict_feat = model.m.predict(np.expand_dims(mfcc, axis=0))
mel_db = predict_feat
output_path = wav[:-4] + '.npy'
np.save(output_path, mel_db)
return centroid
#speakers = ['cl', 'fuli', 'gongwenhua', 'liuyuguang', 'lms', 'lsq', 'lxx', 'lzh', 'shanke', 'wry', 'zhangshuai163', 'zhuting', 'zlb', 'zq', 'yuyaqi']
speakers = ['fuli', 'lzh', 'cl', 'lsq', 'lxx', 'wry', 'lms', 'zlb', 'zq']
enroll_wav_path = 'data_eng'
dict_spkid_embeddings = {}
enroll_nums = 5
total_wavs = glob.glob(os.path.join(data, '*', '*.wav'))
print('total wavs: ', len(total_wavs))
for speaker in speakers:
#print(speaker)
speaker_wavs = glob.glob(os.path.join(enroll_wav_path, speaker, '*.wav'))
length = len(speaker_wavs)
speaker_embeddings = np.zeros((length, 512), dtype=float)
print(speaker, length)
#speaker_embeddings = []
for i in range(length):
#for wav in speaker_wavs:
mfcc = sample_from_mfcc(read_mfcc(speaker_wavs[i], SAMPLE_RATE),
NUM_FRAMES)
predict_feat = model.m.predict(np.expand_dims(mfcc, axis=0))
#speaker_embeddings.append(predict_feat)
speaker_embeddings[i] = predict_feat
dict_spkid_embeddings[speaker] = speaker_embeddings
# num_utterances = len(speaker_wavs)
# enroll_centroid_embeddings = get_centroid(speaker_embeddings, num_utterances)
# dict_spkid_embeddings[speaker] = enroll_centroid_embeddings
# print(speaker, enroll_centroid_embeddings.shape)
np.save('enroll_9.npy', dict_spkid_embeddings)