def time_rnnoise(rounds=1000):
a = rnnoise.RNNoise()
timer = 0.0
st = time.time()
for i in range(rounds):
inp = np.random.bytes(960)
timer = (time.time() - st)
print(timer)
st = time.time()
for i in range(rounds):
inp = np.random.bytes(960)
va, out = a.process_frame(inp)
time_taken_per_frame = ((time.time() - st) - timer) / rounds
print("time taken for one frame - " + str(time_taken_per_frame))
print("time in a frame - " + str(480.0 / 48000.0))
print(str((480.0 / 48000.0) / time_taken_per_frame) + "X faster than real")
a.destroy()
def dnoise(sound):
"""
denoise clip via rnnoise
"""
denoiser = rnnoise.RNNoise()
TARGET_SR = 48000
#audio, sample_rate = rnnoise.read_wave(filename)
sound = sound.set_frame_rate(TARGET_SR)
sound.export('dnntemp.wav', format='wav')
blah = wave.open('dnntemp.wav', 'rb')
blah = blah.readframes(blah.getnframes())
frames = frame_generator(10, blah, TARGET_SR)
frames = list(frames)
tups = [denoiser.process_frame(frame) for frame in frames]
denoised_frames = [tup[1] for tup in tups]
np_audio = np.concatenate(
[np.frombuffer(frame, dtype=np.int16) for frame in denoised_frames])
segment = AudioSegment(data=np_audio.tobytes(),
sample_width=2,
frame_rate=48000,
channels=1)
segment = segment.set_frame_rate(16000)
return segment
def main():
#Globals
global threshold
global input_sample_width
global silence_lenght
global decoder_queue
global segment_lenght
#Setup decoder proc
decoder_proc = multiprocessing.Process(target=process_audio.main,
args=(decoder_queue, ))
decoder_proc.start()
#Noise removal
rnnoise_state = rnnoise.RNNoise()
#Event detection setup
voice_activity = deque(maxlen=int(silence_lenght * audio_sample_density))
slid_win = deque(maxlen=int(silence_lenght * audio_sample_density))
#Place holders
pre_threshold_audio = deque(maxlen=pre_threshold_audio_legnth)
pre_threshold_raw_audio = deque(maxlen=pre_threshold_raw_audio_legnth)
audio_to_send = []
raw_audio_to_send = []
input_data = b''
denoised_data = b''
#Setup loop vars
started = False
started_raw = False
stream = sys.stdin.buffer
print("* Mic set up and listening. ")
try:
while True:
input_data = stream.read(stream_chunk_size)
VodProb, denoised_data = rnnoise_state.process_frame(input_data)
voice_activity.append(VodProb)
#input_data,resampler_state = audioop.ratecv(input_data,input_sample_width,1,input_sample_rate,16000,resampler_state)#If we want to resample audio input
slid_win.append(
math.sqrt(abs(audioop.avg(input_data, input_sample_width))))
#print(max(voice_activity))
#print("-------")
VodProb = max(voice_activity)
#print(max(slid_win))
#threshold_cross = sum([x > threshold for x in slid_win]) > 0
if (VodProb > 0.5):
if started == False:
print("* Starting recording of phrase")
started = True
audio_to_send.append(denoised_data)
pre_threshold_audio, audio_to_send = queAudio(
pre_threshold_audio, audio_to_send, segment_lenght)
elif started:
pre_threshold_audio, audio_to_send = queAudio(
pre_threshold_audio, audio_to_send, 0)
print("* Finished recording, decoding phrase")
#Reset all
started = False
print("* Listening for speech...")
else:
pre_threshold_audio.append(denoised_data)
if (sum([x > threshold for x in slid_win]) > 0):
if started_raw == False:
print("* Starting recording of sound")
started_raw = True
raw_audio_to_send.append(input_data)
raw_pre_threshold_audio, raw_audio_to_send = queAudio(
pre_threshold_raw_audio,
raw_audio_to_send,
segment_lenght,
voice=0)
elif (started_raw):
queAudio(pre_threshold_raw_audio,
raw_audio_to_send,
0,
voice=0)
print("* Finished recording, decoding sound")
#Reset all
started_raw = False
slid_win.clear()
pre_threshold_raw_audio.clear()
raw_audio_to_send = []
print("* Listening ...")
else:
pre_threshold_raw_audio.append(input_data)
except Exception as e:
print(e)
except KeyboardInterrupt:
print("probably ctrl-c")
#exit routine
decoder_proc.terminate()
decoder_proc.join()
rnnoise_state.destroy()
print("* Done listening")
import rnnoise, sys import soundfile as sf denoiser = rnnoise.RNNoise() stream = sys.stdin.buffer input_data = stream.read(480 * 2) va_prob, denoised_data = denoiser.process_frame(input_data)
SPEECH_VOL_MAX = 12
DENOISED_VOL = 3
noise_path = "test/noise"
noise_wavs = getWavs(noise_path)
fp_path = "test/false_samples"
fp_wavs = getWavs(fp_path)
tp_path = "test/true_samples/"
tp_wavs = getWavs(tp_path)
verbose = False
save_true_positives = False
save_false_positives = False
save_true_negatives = False
save_false_negatives = False
rnnoise_state = rnnoise.RNNoise()
def test(x=None):
global DENOISED_VOL
if (x != None):
DENOISED_VOL = x
resampler_state = None
r1 = random.Random()
r2 = random.Random()
r1.seed(83782625373708) #So we get same seq every time
r2.seed(52552468426257) #So we get same seq every time
true_positive_count = 0.0 #Sample had key-phrase and was detected
false_negative_count = 0.0 #Sample had key-phrase but wasn't detected