def test_get_last(self):
buff = CyclicAudioBuffer(3, b'abcdef')
self.assertEqual(buff.get_last(3), b'def')
def test_get_item(self):
buff = CyclicAudioBuffer(6, b'abcdef')
self.assertEqual(buff[:], b'abcdef')
def test_append_with_full(self):
buff = CyclicAudioBuffer(3, b'abc')
buff.append(b'de')
self.assertEqual(buff.get(), b'cde')
self.assertEqual(len(buff), 3)
def test_append_with_room_left(self):
buff = CyclicAudioBuffer(16, b'abc')
buff.append(b'def')
self.assertEqual(buff.get(), b'abcdef')
def test_init_larger_inital_data(self):
size = 16
buff = CyclicAudioBuffer(size, b'a' * (size + 3))
self.assertEqual(buff.get(), b'a' * size)
def test_init(self):
buff = CyclicAudioBuffer(16, b'abc')
self.assertEqual(buff.get(), b'abc')
self.assertEqual(len(buff), 3)
def _wait_until_wake_word(self, source, sec_per_buffer):
"""Listen continuously on source until a wake word is spoken
Arguments:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
"""
# The maximum audio in seconds to keep for transcribing a phrase
# The wake word must fit in this time
ww_duration = self.wake_word_recognizer.expected_duration
ww_test_duration = max(3, ww_duration)
mic_write_counter = 0
num_silent_bytes = int(self.SILENCE_SEC * source.SAMPLE_RATE *
source.SAMPLE_WIDTH)
silence = get_silence(num_silent_bytes)
# Max bytes for byte_data before audio is removed from the front
max_size = source.duration_to_bytes(ww_duration)
test_size = source.duration_to_bytes(ww_test_duration)
audio_buffer = CyclicAudioBuffer(max_size, silence)
buffers_per_check = self.SEC_BETWEEN_WW_CHECKS / sec_per_buffer
buffers_since_check = 0.0
# Rolling buffer to track the audio energy (loudness) heard on
# the source recently. An average audio energy is maintained
# based on these levels.
average_samples = int(5 / sec_per_buffer) # average over last 5 secs
audio_mean = RollingMean(average_samples)
# These are frames immediately after wake word is detected
# that we want to keep to send to STT
ww_frames = deque(maxlen=7)
said_wake_word = False
audio_data = None
while (not said_wake_word and not self._stop_signaled and
not self._skip_wake_word()):
chunk = self.record_sound_chunk(source)
audio_buffer.append(chunk)
ww_frames.append(chunk)
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
audio_mean.append_sample(energy)
if energy < self.energy_threshold * self.multiplier:
self._adjust_threshold(energy, sec_per_buffer)
# maintain the threshold using average
if self.energy_threshold < energy < audio_mean.value * 1.5:
# bump the threshold to just above this value
self.energy_threshold = energy * 1.2
# Periodically output energy level stats. This can be used to
# visualize the microphone input, e.g. a needle on a meter.
if mic_write_counter % 3:
self._watchdog()
self.write_mic_level(energy, source)
mic_write_counter += 1
buffers_since_check += 1.0
# Send chunk to wake_word_recognizer
self.wake_word_recognizer.update(chunk)
if buffers_since_check > buffers_per_check:
buffers_since_check -= buffers_per_check
audio_data = audio_buffer.get_last(test_size) + silence
said_wake_word = \
self.wake_word_recognizer.found_wake_word(audio_data)
self._listen_triggered = False
return WakeWordData(audio_data, said_wake_word,
self._stop_signaled, ww_frames)