def preprocess_audio_pair(speech_file_path, noise_file_path, slice_duration_ms,
n_video_slices, video_frame_rate):
print("preprocessing pair: %s, %s" % (speech_file_path, noise_file_path))
speech_signal = AudioSignal.from_wav_file(speech_file_path)
noise_signal = AudioSignal.from_wav_file(noise_file_path)
while noise_signal.get_number_of_samples(
) < speech_signal.get_number_of_samples():
noise_signal = AudioSignal.concat([noise_signal, noise_signal])
noise_signal.truncate(speech_signal.get_number_of_samples())
factor = AudioMixer.snr_factor(speech_signal, noise_signal, snr_db=0)
noise_signal.amplify_by_factor(factor)
mixed_signal = AudioMixer.mix([speech_signal, noise_signal],
mixing_weights=[1, 1])
mixed_spectrograms = preprocess_audio_signal(mixed_signal,
slice_duration_ms,
n_video_slices,
video_frame_rate)
speech_spectrograms = preprocess_audio_signal(speech_signal,
slice_duration_ms,
n_video_slices,
video_frame_rate)
noise_spectrograms = preprocess_audio_signal(noise_signal,
slice_duration_ms,
n_video_slices,
video_frame_rate)
return mixed_spectrograms, speech_spectrograms, noise_spectrograms, mixed_signal
def enhance_speech(speaker_file_path, noise_file_path, speech_prediction_path,
speech_profile):
print("enhancing mix of %s, %s" % (speaker_file_path, noise_file_path))
speaker_source_signal = AudioSignal.from_wav_file(speaker_file_path)
noise_source_signal = AudioSignal.from_wav_file(noise_file_path)
while noise_source_signal.get_number_of_samples(
) < speaker_source_signal.get_number_of_samples():
noise_source_signal = AudioSignal.concat(
[noise_source_signal, noise_source_signal])
noise_source_signal = noise_source_signal.slice(
0, speaker_source_signal.get_number_of_samples())
mixed_signal = AudioMixer.mix([speaker_source_signal, noise_source_signal])
predicted_speech_signal = AudioSignal.from_wav_file(speech_prediction_path)
signals = [mixed_signal, predicted_speech_signal]
max_length = max([signal.get_number_of_samples() for signal in signals])
for signal in signals:
signal.pad_with_zeros(max_length)
mel_converter = MelConverter(mixed_signal.get_sample_rate(),
n_mel_freqs=128,
freq_min_hz=0,
freq_max_hz=4000)
mixed_spectrogram, original_phase = mel_converter.signal_to_mel_spectrogram(
mixed_signal, get_phase=True)
predicted_speech_spectrogram = mel_converter.signal_to_mel_spectrogram(
predicted_speech_signal)
speech_enhancement_mask = np.zeros(shape=mixed_spectrogram.shape)
thresholds = np.zeros(shape=(speech_enhancement_mask.shape[0]))
for f in range(speech_enhancement_mask.shape[0]):
thresholds[f] = np.percentile(speech_profile[f, :], 85)
for f in range(speech_enhancement_mask.shape[0]):
for t in range(speech_enhancement_mask.shape[1]):
if predicted_speech_spectrogram[f, t] > thresholds[f]:
speech_enhancement_mask[f, t] = 1
continue
enhanced_speech_spectrogram = mixed_spectrogram * speech_enhancement_mask
enhanced_speech_signal = mel_converter.reconstruct_signal_from_mel_spectrogram(
enhanced_speech_spectrogram, original_phase)
return mixed_signal, enhanced_speech_signal
def preprocess_audio_sample(audio_file_path, slice_duration_ms=330):
print("preprocessing %s" % audio_file_path)
audio_signal = AudioSignal.from_wav_file(audio_file_path)
mel_converter = MelConverter(audio_signal.get_sample_rate(),
n_mel_freqs=128,
freq_min_hz=0,
freq_max_hz=4000)
new_signal_length = int(
math.ceil(
float(audio_signal.get_number_of_samples()) /
mel_converter.get_hop_length())) * mel_converter.get_hop_length()
audio_signal.pad_with_zeros(new_signal_length)
mel_spectrogram = mel_converter.signal_to_mel_spectrogram(audio_signal)
samples_per_slice = int(
(float(slice_duration_ms) / 1000) * audio_signal.get_sample_rate())
spectrogram_samples_per_slice = int(samples_per_slice /
mel_converter.get_hop_length())
n_slices = int(mel_spectrogram.shape[1] / spectrogram_samples_per_slice)
slices = [
mel_spectrogram[:, (i * spectrogram_samples_per_slice):(
(i + 1) * spectrogram_samples_per_slice)].flatten()
for i in range(n_slices)
]
return np.stack(slices)
def evaluate(source_file_paths, estimated_file_paths): source_signals = [AudioSignal.from_wav_file(f) for f in source_file_paths] estimated_signals = [AudioSignal.from_wav_file(f) for f in estimated_file_paths] signals = source_signals + estimated_signals max_length = max([signal.get_number_of_samples() for signal in signals]) for signal in signals: signal.pad_with_zeros(max_length) source_data = [signal.get_data(channel_index=0) for signal in source_signals] estimated_data = [signal.get_data(channel_index=0) for signal in estimated_signals] source_data = np.stack(source_data) estimated_data = np.stack(estimated_data) return mir_eval.separation.bss_eval_sources(source_data, estimated_data, compute_permutation=True)
def evaluate(enhancement_dir):
noisy_snr_dbs = []
snr_dbs = []
speaker_ids = os.listdir(enhancement_dir)
for speaker_id in speaker_ids:
for sample_dir_name in os.listdir(
os.path.join(enhancement_dir, speaker_id)):
print('evaluating snr of %s' % sample_dir_name)
source_path = os.path.join(enhancement_dir, speaker_id,
sample_dir_name, 'source.wav')
mixture_path = os.path.join(enhancement_dir, speaker_id,
sample_dir_name, 'mixture.wav')
enhanced_path = os.path.join(enhancement_dir, speaker_id,
sample_dir_name, 'enhanced.wav')
source_signal = AudioSignal.from_wav_file(source_path)
mixture_signal = AudioSignal.from_wav_file(mixture_path)
enhanced_signal = AudioSignal.from_wav_file(enhanced_path)
truncate_longer_signal(mixture_signal, source_signal)
s = source_signal.get_data()
n = mixture_signal.get_data() - source_signal.get_data()
noisy_snr = np.var(s) / np.var(n)
noisy_snr_db = 10 * np.log10(noisy_snr)
print('noisy snr db: %f' % noisy_snr_db)
noisy_snr_dbs.append(noisy_snr_db)
truncate_longer_signal(enhanced_signal, source_signal)
s = source_signal.get_data()
e = enhanced_signal.get_data()
residual_noise = e - s
snr = np.var(s) / np.var(residual_noise)
snr_db = 10 * np.log10(snr)
print('snr db: %f' % snr_db)
snr_dbs.append(snr_db)
print('mean noisy snr db: %f' % np.mean(noisy_snr_dbs))
print('mean snr db: %f' % np.mean(snr_dbs))
def start(args):
# Initialize Network
assets = AssetManager(args.prediction_dir)
storage = PredictionStorage(args.prediction_dir)
network = SpeechEnhancementNetwork.load(
assets.get_model_cache_path(args.model_dir))
network.start_prediction_mode()
network.predict(np.zeros((2, 80, 24)), np.zeros((2, 128, 128, 6)))
predicted_speech_signal = reconstruct_speech_signal\
(AudioSignal.from_wav_file("/cs/engproj/322/real_time/raw_data/mixture.wav"), np.zeros((2, 80, 24)), 30)
with open(assets.get_normalization_cache_path(args.model_dir),
'rb') as normalization_fd:
video_normalizer = pickle.load(normalization_fd)
lock = Lock()
video_dir = assets.get_video_cache_path(args.video_audio_dir)
predict_object = RunPredict(network, video_dir, storage.storage_dir)
# Run video, audio, preprocess and play threads
video_queue = Queue()
audio_queue = Queue()
predict_queue = Queue()
play_queue = Queue()
video_object = VideoProcess(video_dir)
video_thread = Process(target=video_object.capture_frames,
args=(video_queue, lock))
audio_object = AudioProcess(
assets.get_audio_cache_path(args.video_audio_dir))
audio_thread = Process(target=audio_object.capture_frames,
args=(audio_queue, lock))
preprocess_thread = Process(target=predict_object.run_pre_process,
args=(video_queue, audio_queue, predict_queue,
video_normalizer, lock))
play_thread = Process(target=predict_object.play, args=(play_queue, lock))
video_thread.start()
audio_thread.start()
preprocess_thread.start()
play_thread.start()
# Run predict
predict_object.predict(predict_queue, play_queue, lock)
video_thread.join()
audio_thread.join()
preprocess_thread.join()
play_thread.join()
# Save files
predict_object.save_files(storage)
print("*Finish All*")
def build_speech_profile(speaker_speech_dir, max_files=50):
print("building speech profile...")
speech_file_paths = [os.path.join(speaker_speech_dir, f) for f in os.listdir(speaker_speech_dir)][:max_files]
speech_signals = [AudioSignal.from_wav_file(f) for f in speech_file_paths]
mel_converter = MelConverter(speech_signals[0].get_sample_rate(), n_mel_freqs=128, freq_min_hz=0, freq_max_hz=None)
speech_spectrograms = [mel_converter.signal_to_mel_spectrogram(signal) for signal in speech_signals]
speech_profile = np.concatenate(speech_spectrograms, axis=1)
return speech_profile
def separate_sources(source_file_paths, prediction_file_paths,
separation_function):
print("separating mixture of %s" % str(source_file_paths))
source_signals = [AudioSignal.from_wav_file(f) for f in source_file_paths]
prediction_signals = [
AudioSignal.from_wav_file(f) for f in prediction_file_paths
]
signals = source_signals + prediction_signals
max_length = max([signal.get_number_of_samples() for signal in signals])
for signal in signals:
signal.pad_with_zeros(max_length)
mixed_signal = AudioMixer.mix(source_signals)
mel_converter = MelConverter(mixed_signal.get_sample_rate(),
n_mel_freqs=128,
freq_min_hz=0,
freq_max_hz=None)
mixed_spectrogram, original_phase = mel_converter.signal_to_mel_spectrogram(
mixed_signal, get_phase=True)
prediction_spectrograms = [
mel_converter.signal_to_mel_spectrogram(signal)
for signal in prediction_signals
]
masks = generate_separation_masks(mixed_spectrogram,
prediction_spectrograms,
separation_function)
separated_spectrograms = [mixed_spectrogram * mask for mask in masks]
separated_signals = [
mel_converter.reconstruct_signal_from_mel_spectrogram(
s, original_phase) for s in separated_spectrograms
]
return mixed_signal, separated_signals
def reconstruct_signal_from_spectrogram(
magnitude, phase, sample_rate, n_fft, hop_length, mel=True, db=True
):
if db:
magnitude = librosa.db_to_amplitude(magnitude)
if mel:
mel_filterbank = librosa.filters.mel(
sr=sample_rate, n_fft=n_fft, n_mels=80, fmin=0, fmax=8000
)
magnitude = np.dot(np.linalg.pinv(mel_filterbank), magnitude)
signal = librosa.istft(magnitude * phase, hop_length=hop_length)
return AudioSignal(signal, sample_rate)
def reconstruct_signal_from_mel_spectrogram(self,
mel_spectrogram,
log=True,
phase=None):
if log:
mel_spectrogram = librosa.db_to_power(mel_spectrogram)
mel_spectrogram = mel_spectrogram**0.5
magnitude = np.dot(np.linalg.pinv(self._MEL_FILTER), mel_spectrogram)
if phase is not None:
inverted_signal = librosa.istft(magnitude * phase,
hop_length=self._HOP_LENGTH)
else:
inverted_signal = griffin_lim(magnitude,
self._N_FFT,
self._HOP_LENGTH,
n_iterations=10)
return AudioSignal(inverted_signal, self._SAMPLE_RATE)
def run_pre_process(self, v_queue, a_queue, predict_queue,
video_normalizer, lock):
with lock:
print("*Start pre-process*\n")
video_out = cv2.VideoWriter(self.path_video_writer_path,
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'),
self.fps, self.frame_size)
while self.open:
video_frames_list, video_slice_number = v_queue.get()
audio_frames_list, audio_slice_number = a_queue.get()
# Video pre-process
for frame in video_frames_list:
im_crop = self.face_detector.crop_mouth(
frame, self.bounding_box)
video_out.write(im_crop)
im_gray = cv2.cvtColor(im_crop, cv2.COLOR_BGR2GRAY)
self.slice_of_frames[:, :, self.frames_counter] = im_gray
self.frames_counter += 1
slices = [
self.slice_of_frames[:, :, (i * RunPredict.FRAMES_PER_SLICE):(
(i + 1) * RunPredict.FRAMES_PER_SLICE)]
for i in range(RunPredict.NUMBER_OF_SLICES)
]
slices = np.stack(slices)
video_normalizer.normalize(slices)
# Audio pre-process
data = np.fromstring(audio_frames_list, dtype=np.int16)
mixed_signal = AudioSignal(data, 16000)
self.num_iteration += 1
mixed_spectrograms = preprocess_audio_signal(
mixed_signal, self.slice_duration_ms, self.n_video_slices,
self.fps)
# Predict
predict_queue.put(
(slices, mixed_signal, mixed_spectrograms,
int(video_slice_number), int(audio_slice_number)))
self.slice_of_frames = \
np.zeros(shape=(128, 128, RunPredict.NUMBER_OF_FRAMES), dtype=np.float32)
self.frames_counter = 0
if (v_queue.empty() and a_queue.empty()) or audio_slice_number == 0\
or video_slice_number == 0:
with lock:
print(
"****************************************************************"
)
print("Video - slice number: " + str(video_slice_number))
print("Audio - slice number: " + str(audio_slice_number))
print("Predict - number of iterations: " +
str(self.num_iteration))
print(
"****************************************************************"
)
if not v_queue.empty():
video_slice_number -= 1
v_queue.get()
elif not a_queue.empty():
audio_slice_number -= 1
a_queue.get()
v_queue.close()
a_queue.close()
with lock:
print("*Video queue and Audio queue are empty*\n")
break
def preprocess_sample(speech_entry, noise_file_path, slice_duration_ms=200):
print("preprocessing sample: %s, %s, %s..." %
(speech_entry.video_path, speech_entry.audio_path, noise_file_path))
mouth_height = 128
mouth_width = 128
print("preprocessing %s" % speech_entry.video_path)
face_detector = FaceDetector()
a = speech_entry.video_path
with VideoFileReader(a) as reader:
frames = reader.read_all_frames(convert_to_gray_scale=True)
mouth_cropped_frames = np.zeros(shape=(mouth_height, mouth_width, 75),
dtype=np.float32)
for i in range(75):
mouth_cropped_frames[:, :, i] = face_detector.crop_mouth(
frames[i], bounding_box_shape=(mouth_width, mouth_height))
frames_per_slice = int(slice_duration_ms / 1000 *
reader.get_frame_rate())
slices = [
mouth_cropped_frames[:, :,
(i * frames_per_slice):((i + 1) *
frames_per_slice)]
for i in range(int(75 / frames_per_slice))
]
video_samples = np.stack(slices)
video_frame_rate = reader.get_frame_rate()
print("preprocessing pair: %s, %s" %
(speech_entry.audio_path, noise_file_path))
speech_signal = AudioSignal.from_wav_file(speech_entry.audio_path)
print(noise_file_path)
noise_signal = AudioSignal.from_wav_file(noise_file_path)
print(noise_signal.get_data())
print(noise_signal.get_sample_rate())
noise_signal.save_to_wav_file('./noise.wav')
while noise_signal.get_number_of_samples(
) < speech_signal.get_number_of_samples():
noise_signal = AudioSignal.concat([noise_signal, noise_signal])
noise_signal.truncate(speech_signal.get_number_of_samples())
factor = AudioMixer.snr_factor(speech_signal, noise_signal, snr_db=0)
# print(factor)
noise_signal.amplify_by_factor(factor)
#noise_signal.save_to_wav_file('./noise.wav')
mixed_signal = AudioMixer.mix([speech_signal, noise_signal],
mixing_weights=[1, 1])
mixed_signal.save_to_wav_file('./mixed.wav')
mixed_spectrograms = preprocess_audio_signal(mixed_signal,
slice_duration_ms,
video_samples.shape[0],
video_frame_rate)
speech_spectrograms = preprocess_audio_signal(speech_signal,
slice_duration_ms,
video_samples.shape[0],
video_frame_rate)
noise_spectrograms = preprocess_audio_signal(noise_signal,
slice_duration_ms,
video_samples.shape[0],
video_frame_rate)
n_slices = min(video_samples.shape[0], mixed_spectrograms.shape[0])
return Sample(speaker_id=speech_entry.speaker_id,
video_file_path=speech_entry.video_path,
speech_file_path=speech_entry.audio_path,
noise_file_path=noise_file_path,
video_samples=video_samples[:n_slices],
mixed_spectrograms=mixed_spectrograms[:n_slices],
speech_spectrograms=speech_spectrograms[:n_slices],
noise_spectrograms=noise_spectrograms[:n_slices],
mixed_signal=mixed_signal,
video_frame_rate=video_frame_rate)
