def test_wait_while_speaking(self):
# Check that test terminates
create_signal('isSpeaking')
Thread(target=wait_while_speaking_thread).start()
sleep(2)
self.assertFalse(done_waiting)
check_for_signal('isSpeaking')
sleep(2)
self.assertTrue(done_waiting)
def test_check_signal(self):
if exists('/tmp/mycroft'):
rmtree('/tmp/mycroft')
# check that signal is not found if file does not exist
self.assertFalse(check_for_signal('test_signal'))
# Check that the signal is found when created
create_signal('test_signal')
self.assertTrue(check_for_signal('test_signal'))
# Check that the signal is removed after use
self.assertFalse(isfile('/tmp/mycroft/ipc/signal/test_signal'))
def end_audio(self):
"""
Helper function for child classes to call in execute().
Sends the recognizer_loop:audio_output_end message, indicating
that speaking is done for the moment. It also checks if cache
directory needs cleaning to free up disk space.
"""
self.bus.emit(Message("recognizer_loop:audio_output_end"))
# Clean the cache as needed
cache_dir = mycroft.util.get_cache_directory("tts/" + self.tts_name)
mycroft.util.curate_cache(cache_dir, min_free_percent=100)
# This check will clear the "signal"
check_for_signal("isSpeaking")
def main():
""" Main function. Run when file is invoked. """
reset_sigint_handler()
check_for_signal("isSpeaking")
bus = WebsocketClient() # Connect to the Mycroft Messagebus
Configuration.init(bus)
speech.init(bus)
LOG.info("Starting Audio Services")
bus.on('message', create_echo_function('AUDIO', ['mycroft.audio.service']))
audio = AudioService(bus) # Connect audio service instance to message bus
create_daemon(bus.run_forever)
wait_for_exit_signal()
speech.shutdown()
audio.shutdown()
def handle_stop(event):
"""
handle stop message
"""
global _last_stop_signal
if check_for_signal("isSpeaking", -1):
_last_stop_signal = time.time()
tts.playback.clear() # Clear here to get instant stop
bus.emit(Message("mycroft.stop.handled", {"by": "TTS"}))
def handle_stop(event):
"""
handle stop message
"""
global _last_stop_signal
if check_for_signal("isSpeaking", -1):
_last_stop_signal = time.time()
tts.playback.clear_queue()
tts.playback.clear_visimes()
def _skip_wake_word(self):
# Check if told programatically to skip the wake word, like
# when we are in a dialog with the user.
if check_for_signal('startListening'):
return True
# Pressing the Mark 1 button can start recording (unless
# it is being used to mean 'stop' instead)
if check_for_signal('buttonPress', 1):
# give other processes time to consume this signal if
# it was meant to be a 'stop'
sleep(0.25)
if check_for_signal('buttonPress'):
# Signal is still here, assume it was intended to
# begin recording
logger.debug("Button Pressed, wakeword not needed")
return True
return False
def handle_stop(event):
"""
handle stop message
"""
global _last_stop_signal
if check_for_signal("isSpeaking", -1):
_last_stop_signal = time.time()
tts.playback.clear_queue()
tts.playback.clear_visimes()
bus.emit(Message("mycroft.stop.handled", {"by": "TTS"}))
def handle_speak(event):
"""
Handle "speak" message
"""
config = Configuration.get()
Configuration.init(bus)
global _last_stop_signal
# Get conversation ID
if event.context and 'ident' in event.context:
ident = event.context['ident']
else:
ident = 'unknown'
start = time.time() # Time of speech request
with lock:
stopwatch = Stopwatch()
stopwatch.start()
utterance = event.data['utterance']
if event.data.get('expect_response', False):
# When expect_response is requested, the listener will be restarted
# at the end of the next bit of spoken audio.
bus.once('recognizer_loop:audio_output_end', _start_listener)
# This is a bit of a hack for Picroft. The analog audio on a Pi blocks
# for 30 seconds fairly often, so we don't want to break on periods
# (decreasing the chance of encountering the block). But we will
# keep the split for non-Picroft installs since it give user feedback
# faster on longer phrases.
#
# TODO: Remove or make an option? This is really a hack, anyway,
# so we likely will want to get rid of this when not running on Mimic
if (config.get('enclosure', {}).get('platform') != "picroft" and
len(re.findall('<[^>]*>', utterance)) == 0):
chunks = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\;|\?)\s',
utterance)
for chunk in chunks:
# Check if somthing has aborted the speech
if (_last_stop_signal > start or
check_for_signal('buttonPress')):
# Clear any newly queued speech
tts.playback.clear()
break
try:
mute_and_speak(chunk, ident)
except KeyboardInterrupt:
raise
except Exception:
LOG.error('Error in mute_and_speak', exc_info=True)
else:
mute_and_speak(utterance, ident)
stopwatch.stop()
report_timing(ident, 'speech', stopwatch, {'utterance': utterance,
'tts': tts.__class__.__name__})
def handle_speak(event):
global _last_stop_signal
# Mild abuse of the signal system to allow other processes to detect
# when TTS is happening. See mycroft.util.is_speaking()
create_signal("isSpeaking")
utterance = event.data['utterance']
expect_response = event.data.get('expect_response', False)
# This is a bit of a hack for Picroft. The analog audio on a Pi blocks
# for 30 seconds fairly often, so we don't want to break on periods
# (decreasing the chance of encountering the block). But we will
# keep the split for non-Picroft installs since it give user feedback
# faster on longer phrases.
#
# TODO: Remove or make an option? This is really a hack, anyway,
# so we likely will want to get rid of this when not running on Mimic
if not config.get('enclosure', {}).get('platform') == "picroft":
start = time.time()
chunks = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s',
utterance)
for chunk in chunks:
try:
mute_and_speak(chunk)
except KeyboardInterrupt:
raise
except:
logger.error('Error in mute_and_speak', exc_info=True)
if _last_stop_signal > start or check_for_signal('buttonPress'):
break
else:
mute_and_speak(utterance)
# This check will clear the "signal"
check_for_signal("isSpeaking")
if expect_response:
create_signal('startListening')
def show_visimes(self, pairs):
"""
Send visime data to enclosure
Args:
pairs(list): Visime and timing pair
Returns:
True if button has been pressed.
"""
start = time()
for code, duration in pairs:
if check_for_signal('stoppingTTS', -1):
return True
if check_for_signal('buttonPress'):
return True
if self.enclosure:
self.enclosure.mouth_viseme(code)
delta = time() - start
if delta < duration:
sleep(duration - delta)
return False
def visime(self, output):
start = time()
pairs = output.split(" ")
for pair in pairs:
if check_for_signal('buttonPress'):
return
pho_dur = pair.split(":") # phoneme:duration
if len(pho_dur) == 2:
code = VISIMES.get(pho_dur[0], '4')
self.enclosure.mouth_viseme(code)
duration = float(pho_dur[1])
delta = time() - start
if delta < duration:
sleep(duration - delta)
def wait_until_wake_word(self, source, sec_per_buffer):
num_silent_bytes = int(self.SILENCE_SEC * source.SAMPLE_RATE *
source.SAMPLE_WIDTH)
silence = '\0' * num_silent_bytes
# bytearray to store audio in
byte_data = silence
buffers_per_check = self.SEC_BETWEEN_WW_CHECKS / sec_per_buffer
buffers_since_check = 0.0
# Max bytes for byte_data before audio is removed from the front
max_size = self.sec_to_bytes(self.SAVED_WW_SEC, source)
said_wake_word = False
while not said_wake_word:
if check_for_signal('buttonPress'):
said_wake_word = True
continue
chunk = self.record_sound_chunk(source)
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
if energy < self.energy_threshold * self.multiplier:
self.adjust_threshold(energy, sec_per_buffer)
needs_to_grow = len(byte_data) < max_size
if needs_to_grow:
byte_data += chunk
else: # Remove beginning of audio and add new chunk to end
byte_data = byte_data[len(chunk):] + chunk
buffers_since_check += 1.0
if buffers_since_check < buffers_per_check:
buffers_since_check -= buffers_per_check
said_wake_word = self.wake_word_in_audio(byte_data + silence)
def _record_phrase(self, source, sec_per_buffer):
"""Record an entire spoken phrase.
Essentially, this code waits for a period of silence and then returns
the audio. If silence isn't detected, it will terminate and return
a buffer of RECORDING_TIMEOUT duration.
Args:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
Returns:
bytearray: complete audio buffer recorded, including any
silence at the end of the user's utterance
"""
num_loud_chunks = 0
noise = 0
max_noise = 25
min_noise = 0
silence_duration = 0
def increase_noise(level):
if level < max_noise:
return level + 200 * sec_per_buffer
return level
def decrease_noise(level):
if level > min_noise:
return level - 100 * sec_per_buffer
return level
# Smallest number of loud chunks required to return
min_loud_chunks = int(self.MIN_LOUD_SEC_PER_PHRASE / sec_per_buffer)
# Maximum number of chunks to record before timing out
max_chunks = int(self.RECORDING_TIMEOUT / sec_per_buffer)
num_chunks = 0
# Will return if exceeded this even if there's not enough loud chunks
max_chunks_of_silence = int(self.RECORDING_TIMEOUT_WITH_SILENCE /
sec_per_buffer)
# bytearray to store audio in
byte_data = '\0' * source.SAMPLE_WIDTH
phrase_complete = False
while num_chunks < max_chunks and not phrase_complete:
chunk = self.record_sound_chunk(source)
byte_data += chunk
num_chunks += 1
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
test_threshold = self.energy_threshold * self.multiplier
is_loud = energy > test_threshold
if is_loud:
noise = increase_noise(noise)
num_loud_chunks += 1
else:
noise = decrease_noise(noise)
self._adjust_threshold(energy, sec_per_buffer)
if num_chunks % 10 == 0:
with open(self.mic_level_file, 'w') as f:
f.write("Energy: cur=" + str(energy) + " thresh=" +
str(self.energy_threshold))
f.close()
was_loud_enough = num_loud_chunks > min_loud_chunks
quiet_enough = noise <= min_noise
if quiet_enough:
silence_duration += sec_per_buffer
if silence_duration < self.MIN_SILENCE_AT_END:
quiet_enough = False # gotta be silent for min of 1/4 sec
else:
silence_duration = 0
recorded_too_much_silence = num_chunks > max_chunks_of_silence
if quiet_enough and (was_loud_enough or recorded_too_much_silence):
phrase_complete = True
# Pressing top-button will end recording immediately
if check_for_signal('buttonPress'):
phrase_complete = True
return byte_data
def _record_phrase(self, source, sec_per_buffer):
"""Record an entire spoken phrase.
Essentially, this code waits for a period of silence and then returns
the audio. If silence isn't detected, it will terminate and return
a buffer of RECORDING_TIMEOUT duration.
Args:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
Returns:
bytearray: complete audio buffer recorded, including any
silence at the end of the user's utterance
"""
num_loud_chunks = 0
noise = 0
max_noise = 25
min_noise = 0
silence_duration = 0
def increase_noise(level):
if level < max_noise:
return level + 200 * sec_per_buffer
return level
def decrease_noise(level):
if level > min_noise:
return level - 100 * sec_per_buffer
return level
# Smallest number of loud chunks required to return
min_loud_chunks = int(self.MIN_LOUD_SEC_PER_PHRASE / sec_per_buffer)
# Maximum number of chunks to record before timing out
max_chunks = int(self.RECORDING_TIMEOUT / sec_per_buffer)
num_chunks = 0
# Will return if exceeded this even if there's not enough loud chunks
max_chunks_of_silence = int(self.RECORDING_TIMEOUT_WITH_SILENCE /
sec_per_buffer)
# bytearray to store audio in
byte_data = get_silence(source.SAMPLE_WIDTH)
phrase_complete = False
while num_chunks < max_chunks and not phrase_complete:
chunk = self.record_sound_chunk(source)
byte_data += chunk
num_chunks += 1
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
test_threshold = self.energy_threshold * self.multiplier
is_loud = energy > test_threshold
if is_loud:
noise = increase_noise(noise)
num_loud_chunks += 1
else:
noise = decrease_noise(noise)
self._adjust_threshold(energy, sec_per_buffer)
if num_chunks % 10 == 0:
with open(self.mic_level_file, 'w') as f:
f.write("Energy: cur=" + str(energy) + " thresh=" +
str(self.energy_threshold))
f.close()
was_loud_enough = num_loud_chunks > min_loud_chunks
quiet_enough = noise <= min_noise
if quiet_enough:
silence_duration += sec_per_buffer
if silence_duration < self.MIN_SILENCE_AT_END:
quiet_enough = False # gotta be silent for min of 1/4 sec
else:
silence_duration = 0
recorded_too_much_silence = num_chunks > max_chunks_of_silence
if quiet_enough and (was_loud_enough or recorded_too_much_silence):
phrase_complete = True
# Pressing top-button will end recording immediately
if check_for_signal('buttonPress'):
phrase_complete = True
return byte_data
def handle_speak(event):
"""Handle "speak" message
Parse sentences and invoke text to speech service.
"""
config = Configuration.get()
Configuration.set_config_update_handlers(bus)
global _last_stop_signal
# if the message is targeted and audio is not the target don't
# don't synthezise speech
event.context = event.context or {}
if event.context.get('destination') and not \
('debug_cli' in event.context['destination'] or
'audio' in event.context['destination']):
return
# Get conversation ID
if event.context and 'ident' in event.context:
ident = event.context['ident']
else:
ident = 'unknown'
start = time.time() # Time of speech request
with lock:
stopwatch = Stopwatch()
stopwatch.start()
play_error = _should_play_error(event.data)
utterance = event.data['utterance']
listen = event.data.get('expect_response', False)
# This is a bit of a hack for Picroft. The analog audio on a Pi blocks
# for 30 seconds fairly often, so we don't want to break on periods
# (decreasing the chance of encountering the block). But we will
# keep the split for non-Picroft installs since it give user feedback
# faster on longer phrases.
#
# TODO: Remove or make an option? This is really a hack, anyway,
# so we likely will want to get rid of this when not running on Mimic
if (config.get('enclosure', {}).get('platform') != "picroft" and len(
re.findall('<[^>]*>', utterance)) == 0) and not play_error:
# Remove any whitespace present after the period,
# if a character (only alpha) ends with a period
# ex: A. Lincoln -> A.Lincoln
# so that we don't split at the period
utterance = re.sub(r'\b([A-za-z][\.])(\s+)', r'\g<1>', utterance)
chunks = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\;|\?)\s',
utterance)
# Apply the listen flag to the last chunk, set the rest to False
chunks = [(chunks[i], listen if i == len(chunks) - 1 else False)
for i in range(len(chunks))]
for chunk, listen in chunks:
# Check if somthing has aborted the speech
if (_last_stop_signal > start
or check_for_signal('buttonPress')):
# Clear any newly queued speech
tts.playback.clear()
break
try:
mute_and_speak(chunk, ident, listen)
except KeyboardInterrupt:
raise
except Exception:
LOG.error('Error in mute_and_speak', exc_info=True)
else:
mute_and_speak(utterance, ident, listen, play_error)
stopwatch.stop()
report_timing(ident, 'speech', stopwatch, {
'utterance': utterance,
'tts': tts.__class__.__name__
})
def on_stop_handled(self, event):
# A skill performed a stop
check_for_signal('buttonPress')
def record_phrase(self, source, sec_per_buffer):
"""
This attempts to record an entire spoken phrase. Essentially,
this waits for a period of silence and then returns the audio
:rtype: bytearray
:param source: AudioSource
:param sec_per_buffer: Based on source.SAMPLE_RATE
:return: bytearray representing the frame_data of the recorded phrase
"""
num_loud_chunks = 0
noise = 0
max_noise = 25
min_noise = 0
def increase_noise(level):
if level < max_noise:
return level + 200 * sec_per_buffer
return level
def decrease_noise(level):
if level > min_noise:
return level - 100 * sec_per_buffer
return level
# Smallest number of loud chunks required to return
min_loud_chunks = int(self.MIN_LOUD_SEC_PER_PHRASE / sec_per_buffer)
# Maximum number of chunks to record before timing out
max_chunks = int(self.RECORDING_TIMEOUT / sec_per_buffer)
num_chunks = 0
# Will return if exceeded this even if there's not enough loud chunks
max_chunks_of_silence = int(self.RECORDING_TIMEOUT_WITH_SILENCE /
sec_per_buffer)
# bytearray to store audio in
byte_data = '\0' * source.SAMPLE_WIDTH
phrase_complete = False
while num_chunks < max_chunks and not phrase_complete:
chunk = self.record_sound_chunk(source)
byte_data += chunk
num_chunks += 1
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
test_threshold = self.energy_threshold * self.threshold_multiplier
is_loud = energy > test_threshold
if is_loud:
noise = increase_noise(noise)
num_loud_chunks += 1
else:
noise = decrease_noise(noise)
self.adjust_threshold(energy, sec_per_buffer)
was_loud_enough = num_loud_chunks > min_loud_chunks
quiet_enough = noise <= min_noise
recorded_too_much_silence = num_chunks > max_chunks_of_silence
if quiet_enough and (was_loud_enough or recorded_too_much_silence):
phrase_complete = True
if check_for_signal('buttonPress'):
phrase_complete = True
return byte_data
def handle_speak(event):
"""Handle "speak" message
Parse sentences and invoke text to speech service.
"""
#/home/insitelabdev/mycroft-core/mycroft/client/speech/set_config.txt
path = pathlib.Path().absolute()
settings_file = open(
str(path) + '/mycroft/client/speech/set_config.txt', 'r')
settings_dict = eval(settings_file.read())
config = Configuration.get()
Configuration.set_config_update_handlers(bus)
global _last_stop_signal
# Get conversation ID
if event.context and 'ident' in event.context:
ident = event.context['ident']
else:
ident = 'unknown'
start = time.time() # Time of speech request
with lock:
stopwatch = Stopwatch()
stopwatch.start()
utterance = event.data['utterance']
listen = event.data.get('expect_response', False)
# This is a bit of a hack for Picroft. The analog audio on a Pi blocks
# for 30 seconds fairly often, so we don't want to break on periods
# (decreasing the chance of encountering the block). But we will
# keep the split for non-Picroft installs since it give user feedback
# faster on longer phrases.
#
# TODO: Remove or make an option? This is really a hack, anyway,
# so we likely will want to get rid of this when not running on Mimic
if (config.get('enclosure', {}).get('platform') != "picroft"
and len(re.findall('<[^>]*>', utterance)) == 0):
# Remove any whitespace present after the period,
# if a character (only alpha) ends with a period
# ex: A. Lincoln -> A.Lincoln
# so that we don't split at the period
utterance = re.sub(r'\b([A-za-z][\.])(\s+)', r'\g<1>', utterance)
chunks = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\;|\?)\s',
utterance)
# Apply the listen flag to the last chunk, set the rest to False
chunks = [(chunks[i], listen if i == len(chunks) - 1 else False)
for i in range(len(chunks))]
for chunk, listen in chunks:
# Check if somthing has aborted the speech
if (_last_stop_signal > start
or check_for_signal('buttonPress')):
# Clear any newly queued speech
tts.playback.clear()
break
try:
#ana added this
LOG.info("*******THIS IS THE SPEECH SPEED: " +
settings_dict["rate"])
mute_and_speak(
"<speak><prosody volume= " + settings_dict["volume"] +
" rate= " + settings_dict["rate"] + ">" + chunk +
"</prosody></speak>", ident, listen)
except KeyboardInterrupt:
raise
except Exception:
LOG.error('Error in mute_and_speak', exc_info=True)
else:
mute_and_speak(utterance, ident, listen)
stopwatch.stop()
report_timing(ident, 'speech', stopwatch, {
'utterance': utterance,
'tts': tts.__class__.__name__
})
settings_file.close()
def _record_phrase(self,
source,
sec_per_buffer,
stream=None,
ww_frames=None):
"""Record an entire spoken phrase.
Essentially, this code waits for a period of silence and then returns
the audio. If silence isn't detected, it will terminate and return
a buffer of self.recording_timeout duration.
Args:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
stream (AudioStreamHandler): Stream target that will receive chunks
of the utterance audio while it is
being recorded.
ww_frames (deque): Frames of audio data from the last part of wake
word detection.
Returns:
bytearray: complete audio buffer recorded, including any
silence at the end of the user's utterance
"""
num_loud_chunks = 0
noise = 0
max_noise = 25
min_noise = 0
silence_duration = 0
def increase_noise(level):
if level < max_noise:
return level + 200 * sec_per_buffer
return level
def decrease_noise(level):
if level > min_noise:
return level - 100 * sec_per_buffer
return level
# Smallest number of loud chunks required to return
min_loud_chunks = int(self.MIN_LOUD_SEC_PER_PHRASE / sec_per_buffer)
# Maximum number of chunks to record before timing out
max_chunks = int(self.recording_timeout / sec_per_buffer)
num_chunks = 0
# Will return if exceeded this even if there's not enough loud chunks
max_chunks_of_silence = int(self.recording_timeout_with_silence /
sec_per_buffer)
# bytearray to store audio in
byte_data = get_silence(source.SAMPLE_WIDTH)
if stream:
stream.stream_start()
phrase_complete = False
while num_chunks < max_chunks and not phrase_complete:
if ww_frames:
chunk = ww_frames.popleft()
else:
chunk = self.record_sound_chunk(source)
byte_data += chunk
num_chunks += 1
if stream:
stream.stream_chunk(chunk)
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
test_threshold = self.energy_threshold * self.multiplier
is_loud = energy > test_threshold
if is_loud:
noise = increase_noise(noise)
num_loud_chunks += 1
else:
noise = decrease_noise(noise)
self._adjust_threshold(energy, sec_per_buffer)
if num_chunks % 10 == 0:
self.write_mic_level(energy, source)
was_loud_enough = num_loud_chunks > min_loud_chunks
quiet_enough = noise <= min_noise
if quiet_enough:
silence_duration += sec_per_buffer
if silence_duration < self.MIN_SILENCE_AT_END:
quiet_enough = False # gotta be silent for min of 1/4 sec
else:
silence_duration = 0
recorded_too_much_silence = num_chunks > max_chunks_of_silence
if quiet_enough and (was_loud_enough or recorded_too_much_silence):
phrase_complete = True
# Pressing top-button will end recording immediately
if check_for_signal('buttonPress'):
phrase_complete = True
return byte_data
def test_is_speaking(self):
create_signal('isSpeaking')
self.assertTrue(mycroft.audio.is_speaking())
# Check that the signal hasn't been removed
self.assertTrue(check_for_signal('isSpeaking'))
self.assertFalse(mycroft.audio.is_speaking())
def _record_phrase(self,
source,
sec_per_buffer,
stream=None,
ww_frames=None):
"""Record an entire spoken phrase.
Essentially, this code waits for a period of silence and then returns
the audio. If silence isn't detected, it will terminate and return
a buffer of self.recording_timeout duration.
Args:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
stream (AudioStreamHandler): Stream target that will receive chunks
of the utterance audio while it is
being recorded.
ww_frames (deque): Frames of audio data from the last part of wake
word detection.
Returns:
bytearray: complete audio buffer recorded, including any
silence at the end of the user's utterance
"""
noise_tracker = NoiseTracker(0, 25, sec_per_buffer,
self.MIN_LOUD_SEC_PER_PHRASE,
self.recording_timeout_with_silence)
# Maximum number of chunks to record before timing out
max_chunks = int(self.recording_timeout / sec_per_buffer)
num_chunks = 0
# bytearray to store audio in, initialized with a single sample of
# silence.
byte_data = get_silence(source.SAMPLE_WIDTH)
if stream:
stream.stream_start()
phrase_complete = False
while num_chunks < max_chunks and not phrase_complete:
if ww_frames:
chunk = ww_frames.popleft()
else:
chunk = self.record_sound_chunk(source)
byte_data += chunk
num_chunks += 1
if stream:
stream.stream_chunk(chunk)
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
test_threshold = self.energy_threshold * self.multiplier
is_loud = energy > test_threshold
noise_tracker.update(is_loud)
if not is_loud:
self._adjust_threshold(energy, sec_per_buffer)
# The phrase is complete if the noise_tracker end of sentence
# criteria is met or if the top-button is pressed
phrase_complete = (noise_tracker.recording_complete()
or check_for_signal('buttonPress'))
# Periodically write the energy level to the mic level file.
if num_chunks % 10 == 0:
self._watchdog()
self.write_mic_level(energy, source)
return byte_data
def end_audio(self):
"""Helper function for child classes to call in execute()"""
self.ws.emit(Message("recognizer_loop:audio_output_end"))
# This check will clear the "signal"
check_for_signal("isSpeaking")
def handle_speak(event):
"""
Handle "speak" message
"""
config = Configuration.get()
Configuration.init(bus)
global _last_stop_signal
# Get conversation ID
if event.context and 'ident' in event.context:
ident = event.context['ident']
else:
ident = 'unknown'
start = time.time() # Time of speech request
with lock:
stopwatch = Stopwatch()
stopwatch.start()
utterance = event.data['utterance']
if event.data.get('expect_response', False):
# When expect_response is requested, the listener will be restarted
# at the end of the next bit of spoken audio.
bus.once('recognizer_loop:audio_output_end', _start_listener)
# This is a bit of a hack for Picroft. The analog audio on a Pi blocks
# for 30 seconds fairly often, so we don't want to break on periods
# (decreasing the chance of encountering the block). But we will
# keep the split for non-Picroft installs since it give user feedback
# faster on longer phrases.
#
# TODO: Remove or make an option? This is really a hack, anyway,
# so we likely will want to get rid of this when not running on Mimic
if (config.get('enclosure', {}).get('platform') != "picroft"
and len(re.findall('<[^>]*>', utterance)) == 0):
# Remove any whitespace present after the period,
# if a character (only alpha) ends with a period
# ex: A. Lincoln -> A.Lincoln
# so that we don't split at the period
utterance = re.sub(r'\b([A-za-z][\.])(\s+)', r'\g<1>', utterance)
chunks = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\;|\?)\s',
utterance)
for chunk in chunks:
# Check if somthing has aborted the speech
if (_last_stop_signal > start
or check_for_signal('buttonPress')):
# Clear any newly queued speech
tts.playback.clear()
break
try:
mute_and_speak(chunk, ident)
except KeyboardInterrupt:
raise
except Exception:
LOG.error('Error in mute_and_speak', exc_info=True)
else:
mute_and_speak(utterance, ident)
stopwatch.stop()
report_timing(ident, 'speech', stopwatch, {
'utterance': utterance,
'tts': tts.__class__.__name__
})
def _wait_until_wake_word(self, source, sec_per_buffer):
"""Listen continuously on source until a wake word is spoken
Args:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
"""
num_silent_bytes = int(self.SILENCE_SEC * source.SAMPLE_RATE *
source.SAMPLE_WIDTH)
silence = '\0' * num_silent_bytes
# bytearray to store audio in
byte_data = silence
buffers_per_check = self.SEC_BETWEEN_WW_CHECKS / sec_per_buffer
buffers_since_check = 0.0
# Max bytes for byte_data before audio is removed from the front
max_size = self.sec_to_bytes(self.SAVED_WW_SEC, source)
said_wake_word = False
# Rolling buffer to track the audio energy (loudness) heard on
# the source recently. An average audio energy is maintained
# based on these levels.
energies = []
idx_energy = 0
avg_energy = 0.0
energy_avg_samples = int(5 / sec_per_buffer) # avg over last 5 secs
counter = 0
while not said_wake_word:
if check_for_signal('buttonPress'):
said_wake_word = True
continue
chunk = self.record_sound_chunk(source)
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
if energy < self.energy_threshold * self.multiplier:
self._adjust_threshold(energy, sec_per_buffer)
if len(energies) < energy_avg_samples:
# build the average
energies.append(energy)
avg_energy += float(energy)/energy_avg_samples
else:
# maintain the running average and rolling buffer
avg_energy -= float(energies[idx_energy])/energy_avg_samples
avg_energy += float(energy)/energy_avg_samples
energies[idx_energy] = energy
idx_energy = (idx_energy+1) % energy_avg_samples
# maintain the threshold using average
if energy < avg_energy*1.5:
if energy > self.energy_threshold:
# 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 counter % 3:
with open(self.mic_level_file, 'w') as f:
f.write("Energy: cur=" + str(energy) + " thresh=" +
str(self.energy_threshold))
f.close()
counter += 1
# At first, the buffer is empty and must fill up. After that
# just drop the first chunk bytes to keep it the same size.
needs_to_grow = len(byte_data) < max_size
if needs_to_grow:
byte_data += chunk
else: # Remove beginning of audio and add new chunk to end
byte_data = byte_data[len(chunk):] + chunk
buffers_since_check += 1.0
if buffers_since_check > buffers_per_check:
buffers_since_check -= buffers_per_check
said_wake_word = self.wake_word_in_audio(byte_data + silence)
def record_phrase(self, source, sec_per_buffer):
"""
This attempts to record an entire spoken phrase. Essentially,
this waits for a period of silence and then returns the audio
:rtype: bytearray
:param source: AudioSource
:param sec_per_buffer: Based on source.SAMPLE_RATE
:return: bytearray representing the frame_data of the recorded phrase
"""
num_loud_chunks = 0
noise = 0
max_noise = 25
min_noise = 0
def increase_noise(level):
if level < max_noise:
return level + 200 * sec_per_buffer
return level
def decrease_noise(level):
if level > min_noise:
return level - 100 * sec_per_buffer
return level
# Smallest number of loud chunks required to return
min_loud_chunks = int(self.MIN_LOUD_SEC_PER_PHRASE / sec_per_buffer)
# Maximum number of chunks to record before timing out
max_chunks = int(self.RECORDING_TIMEOUT / sec_per_buffer)
num_chunks = 0
# Will return if exceeded this even if there's not enough loud chunks
max_chunks_of_silence = int(self.RECORDING_TIMEOUT_WITH_SILENCE /
sec_per_buffer)
# bytearray to store audio in
byte_data = '\0' * source.SAMPLE_WIDTH
phrase_complete = False
while num_chunks < max_chunks and not phrase_complete:
chunk = self.record_sound_chunk(source)
byte_data += chunk
num_chunks += 1
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
test_threshold = self.energy_threshold * self.multiplier
is_loud = energy > test_threshold
if is_loud:
noise = increase_noise(noise)
num_loud_chunks += 1
else:
noise = decrease_noise(noise)
self.adjust_threshold(energy, sec_per_buffer)
was_loud_enough = num_loud_chunks > min_loud_chunks
quiet_enough = noise <= min_noise
recorded_too_much_silence = num_chunks > max_chunks_of_silence
if quiet_enough and (was_loud_enough or recorded_too_much_silence):
phrase_complete = True
if check_for_signal('buttonPress'):
phrase_complete = True
return byte_data
def test_is_speaking(self):
create_signal('isSpeaking')
self.assertTrue(mycroft.audio.is_speaking())
# Check that the signal hasn't been removed
self.assertTrue(check_for_signal('isSpeaking'))
self.assertFalse(mycroft.audio.is_speaking())
def on_stop_handled(self, event):
# A skill performed a stop
check_for_signal('buttonPress')
def _wait_until_wake_word(self, source, sec_per_buffer):
"""Listen continuously on source until a wake word is spoken
Args:
source (AudioSource): Source producing the audio chunks
sec_per_buffer (float): Fractional number of seconds in each chunk
"""
num_silent_bytes = int(self.SILENCE_SEC * source.SAMPLE_RATE *
source.SAMPLE_WIDTH)
silence = '\0' * num_silent_bytes
# bytearray to store audio in
byte_data = silence
buffers_per_check = self.SEC_BETWEEN_WW_CHECKS / sec_per_buffer
buffers_since_check = 0.0
# Max bytes for byte_data before audio is removed from the front
max_size = self.sec_to_bytes(self.SAVED_WW_SEC, source)
said_wake_word = False
# Rolling buffer to track the audio energy (loudness) heard on
# the source recently. An average audio energy is maintained
# based on these levels.
energies = []
idx_energy = 0
avg_energy = 0.0
energy_avg_samples = int(5 / sec_per_buffer) # avg over last 5 secs
counter = 0
while not said_wake_word:
if check_for_signal('buttonPress'):
said_wake_word = True
continue
chunk = self.record_sound_chunk(source)
energy = self.calc_energy(chunk, source.SAMPLE_WIDTH)
if energy < self.energy_threshold * self.multiplier:
self._adjust_threshold(energy, sec_per_buffer)
if len(energies) < energy_avg_samples:
# build the average
energies.append(energy)
avg_energy += float(energy) / energy_avg_samples
else:
# maintain the running average and rolling buffer
avg_energy -= float(energies[idx_energy]) / energy_avg_samples
avg_energy += float(energy) / energy_avg_samples
energies[idx_energy] = energy
idx_energy = (idx_energy + 1) % energy_avg_samples
# maintain the threshold using average
if energy < avg_energy * 1.5:
if energy > self.energy_threshold:
# 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 counter % 3:
with open(self.mic_level_file, 'w') as f:
f.write("Energy: cur=" + str(energy) + " thresh=" +
str(self.energy_threshold))
f.close()
counter += 1
# At first, the buffer is empty and must fill up. After that
# just drop the first chunk bytes to keep it the same size.
needs_to_grow = len(byte_data) < max_size
if needs_to_grow:
byte_data += chunk
else: # Remove beginning of audio and add new chunk to end
byte_data = byte_data[len(chunk):] + chunk
buffers_since_check += 1.0
if buffers_since_check > buffers_per_check:
buffers_since_check -= buffers_per_check
said_wake_word = self.wake_word_in_audio(byte_data + silence)
