def load_featurizer_model(self, featurizer_path):
""" load the given compiled ELL featurizer for use in processing subsequent audio input """
if featurizer_path:
self.featurizer = featurizer.AudioTransform(featurizer_path, 40)
self.setup_spectrogram_image()
self.vad = vad.VoiceActivityDetector(self.sample_rate,
self.featurizer.output_size)
self.show_output("Feature input size: {}, output size: {}".format(
self.featurizer.input_size, self.featurizer.output_size))
self.init_data()
def create_vad(self):
vad_options = {
"tau_up": self.get_entry(self.tau_up),
"tau_down": self.get_entry(self.tau_down),
"threshold_up": self.get_entry(self.threshold_up),
"threshold_down": self.get_entry(self.threshold_down),
"large_input": self.get_entry(self.large_input),
"gain_att": self.get_entry(self.gain_att),
"level_threshold": self.get_entry(self.level_threshold)
}
model = make_vad.make_vad("vad.ell", self.sample_rate,
self.featurizer.input_size,
self.featurizer.output_size, vad_options)
self.vad = vad.VoiceActivityDetector(model)
def __init__(self,
featurizer_model=None,
classifier_model=None,
auto_scale=True,
sample_rate=None,
channels=None,
input_device=None,
categories=None,
image_width=80,
threshold=None,
wav_file=None,
clear=5,
serial=None,
vad_model=None,
smoothing=None,
ignore_list=None):
""" Initialize AudioDemo object
featurizer_model - the path to the ELL featurizer
classifier_model - the path to the ELL classifier
auto_scale - auto scale audio input to range [-1, 1]
sample_rate - sample rate to featurizer is expecting
channels - number of channels featurizer is expecting
input_device - optional id of microphone to use
categories - path to file containing category labels
image_width - width of the spectrogram image
threshold - ignore predictions that have confidence below this number (e.g. 0.5)
wav_file - optional wav_file to use when you click Play
serial - optional serial input, reading numbers from the given serial port.
vad_model - optional ELL model containing VoiceActivityDetector
smoothing - controls the size of the smoothing window (defaults to 0).
ignore_list - list of category labels to ignore (like 'background' or 'silence')
"""
super().__init__()
self.CLASSIFIER_MODEL_KEY = "classifier_model"
self.FEATURIZER_MODEL_KEY = "featurizer_model"
self.WAV_FILE_KEY = "wav_file"
self.CATEGORY_FILE_KEY = "categories"
self.get_settings_file_name()
self.load_settings()
self.reading_input = False
self.featurizer_model = None
self.serial_port = serial
self.smoothing = smoothing
self.ignore_list = ignore_list
if featurizer_model:
self.featurizer_model = featurizer_model
self.settings[self.FEATURIZER_MODEL_KEY] = featurizer_model
elif self.FEATURIZER_MODEL_KEY in self.settings:
self.featurizer_model = self.settings[self.FEATURIZER_MODEL_KEY]
self.classifier_model = None
if classifier_model:
self.classifier_model = classifier_model
self.settings[self.CLASSIFIER_MODEL_KEY] = classifier_model
elif self.CLASSIFIER_MODEL_KEY in self.settings:
self.classifier_model = self.settings[self.CLASSIFIER_MODEL_KEY]
self.wav_filename = wav_file
if self.wav_filename is None and self.WAV_FILE_KEY in self.settings:
self.wav_filename = self.settings[self.WAV_FILE_KEY]
self.wav_file_list = None
self.auto_scale = auto_scale
self.sample_rate = sample_rate if sample_rate is not None else 16000
self.channels = channels if channels is not None else 1
self.input_device = input_device
self.num_classifier_features = None
self.vad = None
self.vad_reset = (vad_model is not None)
self.previous_vad = 0
self.vad_latch = 3 # only reset after 3 vad=0 signals to smooth vad signal a bit.
if not categories and self.CATEGORY_FILE_KEY in self.settings:
categories = self.settings[self.CATEGORY_FILE_KEY]
self.categories = categories
if categories:
self.settings[self.CATEGORY_FILE_KEY] = categories
self.save_settings() # in case we just changed it.
self.audio_level = 0
self.min_level = 0
self.max_level = 0
self.threshold = threshold
self.output_clear_time = int(clear * 1000) if clear else 5000
self.featurizer = None
self.classifier = None
self.wav_file = None
self.speaker = None
self.microphone = None
self.animation = None
self.show_classifier_output = True
self.last_prediction = None
self.probability = 0
# Threads
self.read_input_thread = None
self.lock = Lock()
self.main_thread = get_ident()
self.message_queue = []
# UI components
self.max_spectrogram_width = image_width
self.features_entry = None
self.classifier_feature_data = None
self.spectrogram_image_data = None
self.init_ui()
if self.featurizer_model:
self.load_featurizer_model(os.path.abspath(self.featurizer_model))
else:
self.show_output("Please specify and load a feature model")
if smoothing == "vad":
# smooth up to 1 second worth of predictions
self.smoothing = int(self.sample_rate / self.featurizer.input_size)
if vad_model is None:
vad_model = make_vad.make_vad("vad.ell", self.sample_rate,
self.featurizer.input_size,
self.featurizer.output_size,
None)
if self.classifier_model:
self.load_classifier(self.classifier_model)
self.setup_spectrogram_image()
else:
self.show_output("Please specify and load a classifier model")
if vad_model:
self.vad = vad.VoiceActivityDetector(vad_model)