def __call__(self, context: SpeechContext, frame: np.ndarray) -> None:
"""Processes a single frame of audio to determine if voice is present
Args:
context (SpeechContext): State based information that needs to be shared
between pieces of the pipeline
frame (np.ndarray): Single frame of PCM-16 audio from an input source
"""
frame = frame.tobytes()
result: bool = self._vad.is_speech(frame, self._sample_rate)
raw = result > 0
if raw == self._run_value:
self._run_length += 1
else:
self._run_value = raw
self._run_length = 1
if self._run_value != context.is_speech:
if self._run_value and self._run_length >= self._rise_length:
context.is_speech = True
_LOG.info("vad: true")
if not self._run_value and self._run_length >= self._fall_length:
context.is_speech = False
_LOG.info("vad: false")
def __call__(self, context: SpeechContext, frame: np.ndarray) -> None:
"""Processes a single frame of audio to determine if voice is present
Args:
context (SpeechContext): State based information that needs to be shared
between pieces of the pipeline
frame (np.ndarray): Single frame of PCM-16 audio from an input source
"""
# validate dtype
if not np.issubdtype(frame.dtype, np.signedinteger):
raise TypeError("invalid_dtype")
result: bool = self._vad.is_speech(frame)
raw = result > 0
if raw == self._run_value:
self._run_length += 1
else:
self._run_value = raw
self._run_length = 1
if self._run_value != context.is_speech:
if self._run_value and self._run_length >= self._rise_length:
context.is_speech = True
_LOG.info("vad: true")
if not self._run_value and self._run_length >= self._fall_length:
context.is_speech = False
_LOG.info("vad: false")
def test_detect_activate(_mock):
context = SpeechContext()
detector = WakewordTrigger(model_dir="wakeword_model")
detector.detect_model.return_value[0][:] = 0.6
test_frame = np.random.rand(512, ).astype(np.float32)
context.is_speech = True
detector(context, test_frame)
context.is_speech = False
assert context.is_active
def test_detect_inactive_vad_deactivate(_mock):
context = SpeechContext()
detector = WakewordTrigger(model_dir="wakeword_model")
for _ in range(3):
test_frame = np.random.rand(160, ).astype(np.float32)
context.is_speech = True
detector(context, test_frame)
context.is_speech = False
assert not context.is_active
detector(context, test_frame)
def test_context():
context = SpeechContext()
# test is_speech
assert not context.is_speech
context.is_speech = True
assert context.is_speech
# test is_active
assert not context.is_active
context.is_active = True
assert context.is_active
# test transcript
assert not context.transcript
context.transcript = "this is a test"
assert context.transcript
# test confidence
assert context.confidence == 0.0
context.confidence = 1.0
assert context.confidence == 1.0
# test reset
context.reset()
assert not context.is_speech
assert not context.is_active
assert not context.transcript
assert context.confidence == 0.0
def test_detect_vad_inactive(_mock):
context = SpeechContext()
detector = WakewordTrigger(model_dir="wakeword_model")
test_frame = np.random.rand(160, ).astype(np.float32)
context.is_speech = False
detector(context, test_frame)
assert not context.is_active
def test_min_active():
max_active = 500
min_active = 120
context = SpeechContext()
timeout = ActivationTimeout(min_active=min_active, max_active=max_active)
context.is_active = True
# call with speech active
context.is_speech = True
timeout(context)
# call timeout after speech is no longer detected
context.is_speech = False
timeout(context)
assert context.is_active
# vad fall should be True
# with context still active
timeout(context)
assert context.is_active
# context should remain active until min active
steps_before_deactivate = min_active // 20
for _ in range(steps_before_deactivate):
timeout(context)
assert context.is_active
# call with speech active
context.is_speech = True
timeout(context)
# call timeout after speech is no longer detected
# min active should be satisfied
context.is_speech = False
timeout(context)
assert not context.is_active
timeout.close()
def test_timeout_vad_fall():
max_active = 500
min_active = 20
context = SpeechContext()
timeout = ActivationTimeout(min_active=min_active, max_active=max_active)
context.is_active = True
context.is_speech = False
timeout(context)
context.is_speech = True
timeout(context)
assert context.is_active
context.is_speech = False
steps_before_timeout = (min_active // 20) + 2
for _ in range(steps_before_timeout):
timeout(context)
assert not context.is_active
timeout.close()
def test_voice_activity_trigger():
context = SpeechContext()
trigger = VoiceActivityTrigger()
frame = np.zeros(160, np.int16)
trigger(context, frame)
assert not context.is_active
context.is_speech = True
trigger(context, frame)
assert context.is_active
trigger.close()
