def getModel(ARGS):
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.scorer = os.path.join(model_dir, ARGS.scorer)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
model = deepspeech.Model(ARGS.model)
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.scorer = os.path.join(model_dir, ARGS.scorer)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
model = deepspeech.Model(ARGS.model)
#model.addHotWord('Fire', 10)
#model.addHotWord('Intruder', 10)
#model.addHotWord('Help', 10)
#model.addHotWord('Yes', 10)
#model.addHotWord('No', 10)
if ARGS.scorer:
logging.info("ARGS.scorer: %s", ARGS.scorer)
model.enableExternalScorer(ARGS.scorer)
return model
def _register_(serviceList, pluginProperties):
global services, plugin, core, audioRecorder, dsModel, stream, defaudsrc, actions, replacements, pattern
services = serviceList
plugin = pluginProperties
core = services["core"][0]
audioRecorder = services["audioRecorder"][0]
actions = services["actions"][0]
with open(r"englishSTT\en_us_replacements.json") as f:
replacements = json.load(f)
replacements = dict((re.escape(k), v) for k, v in replacements.items())
pattern = re.compile("|".join(replacements.keys()))
defaudsrc = audioRecorder.getAudioSource(device=1)
try:
dsModel = deepspeech.Model(
r"englishSTT/deepspeech-0.9.3-models/deepspeech-0.9.3-models.pbmm",
)
dsModel.enableExternalScorer(
r"englishSTT/deepspeech-0.9.3-models/deepspeech-0.9.3-models.scorer"
)
except RuntimeError:
print("Downloading deepspeech models")
os.makedirs("englishSTT/deepspeech-0.9.3-models")
print("Downloading deepspeech .pbmm")
urllib.request.urlretrieve(
"https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.pbmm",
"englishSTT/deepspeech-0.9.3-models/deepspeech-0.9.3-models.pbmm")
print("Downloading deepspeech .scorer")
urllib.request.urlretrieve(
"https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.scorer",
"englishSTT/deepspeech-0.9.3-models/deepspeech-0.9.3-models.scorer"
)
dsModel = deepspeech.Model(
r"englishSTT/deepspeech-0.9.3-models/deepspeech-0.9.3-models.pbmm",
)
dsModel.enableExternalScorer(
r"englishSTT/deepspeech-0.9.3-models/deepspeech-0.9.3-models.scorer"
)
#dsModel.enableDecoderWithLM(
# r"englishSTT\model\lm.binary",
# r"englishSTT\model\trie",
# 0.75,
# 1.85)
# DeepSpeech locks up after the first few frames, this clears that up
# THIS WAS ADDED WITH DEEPSPEECH 0.6.0, AND MAY NO LONGER BE NEEDED
stream = dsModel.createStream()
stream.feedAudioContent([0, 0, 0, 0, 65535, 65535, 65535, 65535] * 8192)
stream.finishStream()
services["userInterface"][0].addCommands({"trigger": trigger})
def main():
model = deepspeech.Model(MODEL_PATH)
model.setBeamWidth(BEAM_WIDTH)
model.enableExternalScorer(SCORER_PATH)
stream = model.createStream()
audio = pyaudio.PyAudio()
index, name = find_device(audio, 'pulse')
print(f'select device {name}')
buffer_size = model.sampleRate() // BUFFERS_PER_SECOND
audio_stream = audio.open(rate=model.sampleRate(),
channels=1,
format=audio.get_format_from_width(
SAMPLE_WIDTH, unsigned=False),
input_device_index=index,
input=True,
frames_per_buffer=buffer_size,
stream_callback=audio_callback)
num_iterations = BUFFERS_PER_SECOND * 2
i = 0
while audio_stream.is_active():
stream.feedAudioContent(buffer_queue.get())
if i % num_iterations == 0:
text = stream.intermediateDecode()
if text.find('stop') >= 0:
break
print(text)
i += 1
print(stream.finishStream())
audio_stream.close()
def maybe_load_model(self):
"""Load DeepSpeech model if not already loaded."""
if self.model:
return
assert self.model_path, "No model path"
_LOGGER.debug("Loading model from %s (beam width=%s)", self.model_path,
self.beam_width)
self.model = deepspeech.Model(str(self.model_path), self.beam_width)
if (self.language_model_path and self.language_model_path.is_file()
and self.trie_path and self.trie_path.is_file()):
_LOGGER.debug(
"Enabling language model (lm=%s, trie=%s, lm_alpha=%s, lm_beta=%s)",
self.language_model_path,
self.trie_path,
self.lm_alpha,
self.lm_beta,
)
self.model.enableDecoderWithLM(
str(self.language_model_path),
str(self.trie_path),
self.lm_alpha,
self.lm_beta,
)
async def main():
BEAM_WIDTH = 500
LM_ALPHA = 0.75
LM_BETA = 1.85
MODEL = "./models/deepspeech-0.8.2-models.pbmm"
LANG_MODEL = "./models/lm.binary"
TRIE = "./models/trie"
model = deepspeech.Model(MODEL, BEAM_WIDTH)
model.enableDecoderWithLM(LANG_MODEL, TRIE, LM_ALPHA, LM_BETA)
vad_audio = VADAudio(aggressiveness=3, device=11, input_rate=48000)
uri = "ws://localhost:8000/ws"
ws = await websockets.connect(uri, ping_interval=None)
stream_context = model.createStream()
for frame in vad_audio.vad_collector():
if frame is not None:
model.feedAudioContent(stream_context, np.frombuffer(frame, np.int16))
else:
text = model.finishStream(stream_context)
try:
await ws.send(text)
stream_context = model.createStream()
returned = await ws.recv()
print(returned)
except: # there is clearly a better ay to do this but I have a smol brain
print("Reconnecting")
ws = await websockets.connect(uri, ping_interval=None)
def main(ARGS):
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.alphabet = os.path.join(
model_dir, ARGS.alphabet if ARGS.alphabet else 'alphabet.txt')
ARGS.lm = os.path.join(model_dir, ARGS.lm)
ARGS.trie = os.path.join(model_dir, ARGS.trie)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
logging.info("ARGS.alphabet: %s", ARGS.alphabet)
model = deepspeech.Model(ARGS.model, ARGS.beam_width)
if ARGS.lm and ARGS.trie:
logging.info("ARGS.lm: %s", ARGS.lm)
logging.info("ARGS.trie: %s", ARGS.trie)
model.enableDecoderWithLM(ARGS.lm, ARGS.trie, ARGS.lm_alpha,
ARGS.lm_beta)
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate,
file=ARGS.file)
# this create a generator from vad_collector function.
# if new block is saved it will be saved in the buffer queue
# when a new frame is required to process the function vad_collector.
# provids the next uttrance by using yeild, so it will provide frame and
# when next one is asked it will return to the function from the last yeild postion.
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not ARGS.nospinner:
spinner = Halo(spinner='line')
stream_context = model.createStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
model.feedAudioContent(stream_context,
np.frombuffer(frame, np.int16))
if ARGS.savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS.savewav:
vad_audio.write_wav(
os.path.join(
ARGS.savewav,
datetime.now().strftime(
"savewav_%Y-%m-%d_%H-%M-%S_%f.wav")), wav_data)
wav_data = bytearray()
text = model.finishStream(stream_context)
print("Recognized: %s" % text)
stream_context = model.createStream()
def __init__(self, results_event, config=None):
if len(
signature(super(DeepSpeechLocalStreamingSTT,
self).__init__).parameters) == 2:
super(DeepSpeechLocalStreamingSTT,
self).__init__(results_event, config)
else:
LOG.warning(
f"Shorter Signature Found; config will be ignored and results_event will not be handled!"
)
super(DeepSpeechLocalStreamingSTT, self).__init__()
self.results_event = None
# override language with module specific language selection
self.language = self.config.get('lang') or self.lang
self.queue = None
if not self.language.startswith("en"):
raise ValueError("DeepSpeech is currently english only")
model_path = self.config.get("model_path") or \
os.path.expanduser("~/.local/share/neon/deepspeech-0.9.3-models.pbmm")
scorer_path = self.config.get("scorer_path") or \
os.path.expanduser("~/.local/share/neon/deepspeech-0.9.3-models.scorer")
if not os.path.isfile(model_path):
LOG.error("Model not found and will be downloaded!")
LOG.error(model_path)
get_model()
self.client = deepspeech.Model(model_path)
if not scorer_path or not os.path.isfile(scorer_path):
LOG.warning("You should provide a valid scorer")
LOG.info(
"download scorer from https://github.com/mozilla/DeepSpeech")
else:
self.client.enableExternalScorer(scorer_path)
def __init__(self):
self.BEAM_WIDTH = 500
self.LM_ALPHA = 0.75
self.LM_BETA = 1.85
self.model_dir = 'DeepSpeech/data/wernicke/model/'
self.model_file = os.path.join(self.model_dir, 'output_graph.pb')
# self.model_dir = 'deepspeech-0.6.0-models/'
# self.model_file = os.path.join(self.model_dir, 'output_graph.pbmm')
self.lm_file = os.path.join(self.model_dir, 'lm.binary')
self.trie_file = os.path.join(self.model_dir, 'trie')
self.save_dir = 'saved_wavs'
os.makedirs(self.save_dir, exist_ok=True)
# load segment model
log.info('Initializing pyAudioAnalysis classifier model...')
[
self.classifier, self.MEAN, self.STD, self.class_names,
self.mt_win, self.mt_step, self.st_win, self.st_step, _
] = aT.load_model("wernicke_server_model")
self.fs = 16000
log.info('Initializing deepspeech model...')
self.model = deepspeech.Model(self.model_file, self.BEAM_WIDTH)
# Temporarily disabling this. I don't think I have nearly enough samples to start doing LM and trie files, etc
self.model.enableDecoderWithLM(self.lm_file, self.trie_file,
self.LM_ALPHA, self.LM_BETA)
log.info('Models ready.')
def main():
# setup pre trained model for audio to text transcribing
model_file_path = 'deepspeech-0.6.0-models/output_graph.pbmm'
beam_width = 500
model = deepspeech.Model(model_file_path, beam_width)
lm_file_path = 'deepspeech-0.6.0-models/lm.binary'
trie_file_path = 'deepspeech-0.6.0-models/trie'
lm_alpha = 0.75
lm_beta = 1.85
model.enableDecoderWithLM(lm_file_path, trie_file_path, lm_alpha, lm_beta)
# get MongoDb client
podcast_db = get_db()
podcast_fs = get_fs(podcast_db)
# find all segments that are not transcribed
for segment in podcast_db.segment.find({"segment_transcript": None}, \
no_cursor_timeout=True):
key = segment['gridfs_key']
# read wav audio file for this segment
data = podcast_fs.get(key)
audio = wave.open(data, 'rb')
# transcribe this audio segment to text
transcript = transcribe_audio_to_text(audio, model)
audio.close()
# print(transcript)
segment['segment_transcript'] = transcript
# add updated MongoDb record into the collection
update_record(podcast_db, segment)
def main(ARGS):
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.scorer = os.path.join(model_dir, ARGS.scorer)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
model = deepspeech.Model(ARGS.model)
if ARGS.scorer:
logging.info("ARGS.scorer: %s", ARGS.scorer)
model.enableExternalScorer(ARGS.scorer)
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate,
file=ARGS.file)
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not ARGS.nospinner:
spinner = Halo(spinner='line')
# Deepspeech model
stream_context = model.createStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
stream_context.feedAudioContent(np.frombuffer(frame, np.int16))
if ARGS.savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS.savewav:
vad_audio.write_wav(
os.path.join(
ARGS.savewav,
datetime.now().strftime(
"savewav_%Y-%m-%d_%H-%M-%S_%f.wav")), wav_data)
wav_data = bytearray()
text = stream_context.finishStream()
vad_audio.stream.stop_stream()
print("Recognized: %s" % text)
if text != "":
# send text as a ros topic, due to being in a python3 env not ros, calling command line
subprocess.run([
'rostopic', 'pub', '--once', '/gaan/nlp/user_msg',
'std_msgs/String', text
])
# Listen from microphone
vad_audio.stream.start_stream()
stream_context = model.createStream()
def create_model_from_config(config: Config) -> deepspeech.Model:
print("Initialize model...")
model = deepspeech.Model(config.model_path)
model.setBeamWidth(config.beam_width)
model.enableExternalScorer(config.scorer)
model.setScorerAlphaBeta(config.scorer_alpha, config.scorer_beta)
return model
def __init__(self):
super().__init__()
start_time = time.perf_counter()
LOG.info("Loading DeepSpeech model...")
model = self.config['model']
alphabet = self.config['alphabet']
num_context = self.config.get('num_context', 9)
beam_width = self.config.get('beam_width', 512)
num_features = self.config.get('num_features', 26)
lm = self.config.get('lm')
trie = self.config.get('trie')
self.model = deepspeech.Model(model, num_features, num_context,
alphabet, beam_width)
if lm is not None and trie is not None:
lm_weight = self.config.get('lm_weight', 1.5)
vwcw = self.config.get('valid_word_count_weight', 2.25)
self.model.enableDecoderWithLM(alphabet, lm, trie, lm_weight, vwcw)
LOG.info("Loaded DeepSpeech model in %0.3fs" %
(time.perf_counter() - start_time))
self.stream_ctx = None
self.can_stream = True
def _get_model(self) -> deepspeech.Model:
if not self._model:
self._model = deepspeech.Model(self.model_file, self.beam_width)
self._model.enableDecoderWithLM(self.lm_file, self.trie_file,
self.lm_alpha, self.lm_beta)
return self._model
def __init__(self, results_event, config=None):
super(DeepSpeechLocalStreamingSTT,
self).__init__(results_event, config)
# override language with module specific language selection
self.language = self.config.get('lang') or self.lang
self.queue = None
if not self.language.startswith("en"):
raise ValueError("DeepSpeech is currently english only")
model_path = self.config.get("model_path") or \
os.path.expanduser("~/.local/share/neon/deepspeech-0.8.1-models.pbmm")
scorer_path = self.config.get("scorer_path") or \
os.path.expanduser("~/.local/share/neon/deepspeech-0.8.1-models.scorer")
if not os.path.isfile(model_path):
LOG.error("You need to provide a valid model file")
LOG.error(model_path)
LOG.info(
"download a model from https://github.com/mozilla/DeepSpeech")
raise FileNotFoundError
if not scorer_path or not os.path.isfile(scorer_path):
LOG.warning("You should provide a valid scorer")
LOG.info(
"download scorer from https://github.com/mozilla/DeepSpeech")
self.client = deepspeech.Model(model_path)
if scorer_path:
self.client.enableExternalScorer(scorer_path)
async def activate(self, site):
# self.log('activate')
#if not self.active[site]:
if os.path.isdir(self.model_path):
# self.log('START DS ASR')
self.audio_stream[site] = BytesLoop()
self.active[site] = True
self.started[site] = False
await self.client.subscribe('hermod/' + site + '/microphone/audio')
# Load DeepSpeech model
# self.log('START DS ASR ACTIVATE '+self.model_path)
#deepspeech-0.7.0-models.pbmm
modelPath = os.path.join(self.model_path, self.modelFile)
scorerPath = os.path.join(self.model_path,
'deepspeech-0.7.0-models.scorer')
# lm = os.path.join(self.model_path, 'lm.binary')
# trie = os.path.join(self.model_path, 'trie')
self.log('START DS ASR ACTIVATE ' + modelPath)
# self.models[site] = deepspeech.Model(modelPath, 500)
# if lm and trie:
# self.models[site].enableDecoderWithLM(lm, trie, 0.75, 1.85)
self.models[site] = deepspeech.Model(modelPath)
self.models[site].enableExternalScorer(scorerPath)
self.stream_contexts[site] = self.models[site].createStream()
def deepspeech_stt():
ARGS_model = 'models/deepspeech.pbmm'
ARGS_scorer = 'models/deepspeech.scorer'
model = deepspeech.Model(ARGS_model)
if ARGS_scorer:
model.enableExternalScorer(ARGS_scorer)
vad_audio = VADAudio(aggressiveness=0,
device=None,
input_rate=16000,
file=None)
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
spinner = None
ARGS_savewav = 1
#spinner = Halo(spinner='line')
stream_context = model.createStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
stream_context.feedAudioContent(np.frombuffer(frame, np.int16))
if ARGS_savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS_savewav:
vad_audio.write_wav("input_temp.wav", wav_data)
wav_data = bytearray()
text = stream_context.finishStream()
print("Recognized: %s" % text)
stream_context = model.createStream()
return text
def __init__(self):
dirName = os.path.expanduser('deep_speech_models')
model_path = dirName + '/deepspeech-0.7.0-models.pbmm'
scorer_path = dirName + '/deepspeech-0.7.0-models.scorer'
self.dir_audio = 'audio_tests/'
self.model = deepspeech.Model(model_path)
self.model.enableExternalScorer(scorer_path)
def main(ARGS):
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'custom_lm_output_graph.pb')
ARGS.alphabet = os.path.join(model_dir, ARGS.alphabet if ARGS.alphabet else 'alphabet.txt')
ARGS.lm = os.path.join(model_dir, ARGS.lm)
ARGS.trie = os.path.join(model_dir, ARGS.trie)
print("Booting up server...")
server = Server()
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
logging.info("ARGS.alphabet: %s", ARGS.alphabet)
model = deepspeech.Model(ARGS.model, ARGS.n_features, ARGS.n_context, ARGS.alphabet, ARGS.beam_width)
if ARGS.lm and ARGS.trie:
logging.info("ARGS.lm: %s", ARGS.lm)
logging.info("ARGS.trie: %s", ARGS.trie)
model.enableDecoderWithLM(ARGS.alphabet, ARGS.lm, ARGS.trie, ARGS.lm_alpha, ARGS.lm_beta)
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate)
# Stream from microphone to DeepSpeech using VAD
stream_context = model.setupStream()
wav_data = bytearray()
print("Warming up model...")
# Warm up the model - For some reason there's a few seconds long pause on the 26th frame we feed into the model, as
# it's presumable reallocating memory or something? In any case, if we do this up front latencies are much better.
empty_frame = np.zeros((320,), dtype=np.int16)
for i in range(26):
model.feedAudioContent(stream_context, empty_frame)
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# frames = audio.frame_generator()
count = 0
for frame in frames:
if frame is not None:
model.feedAudioContent(stream_context, np.frombuffer(frame, np.int16))
count += 1
if count > 20:
text = model.intermediateDecode(stream_context)
count = 0
if len(text) > 0:
print("Intermediate recognition: %s" % text)
server.emit_utterance(text, True)
else:
text = model.finishStream(stream_context)
stream_context = model.setupStream()
if len(text) > 0:
print("Recognized: %s" % text)
server.emit_utterance(text, False)
stream_context = model.setupStream()
def main(ARGS):
# init node
pub = rospy.Publisher('chatter', String, queue_size=10)
rospy.init_node('talker', anonymous=True)
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.lm = os.path.join(model_dir, ARGS.lm)
ARGS.trie = os.path.join(model_dir, ARGS.trie)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
model = deepspeech.Model(ARGS.model, ARGS.beam_width)
if ARGS.lm and ARGS.trie:
logging.info("ARGS.lm: %s", ARGS.lm)
logging.info("ARGS.trie: %s", ARGS.trie)
model.enableDecoderWithLM(ARGS.lm, ARGS.trie, ARGS.lm_alpha,
ARGS.lm_beta)
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate,
file=ARGS.file)
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not ARGS.nospinner:
spinner = Halo(spinner='line')
stream_context = model.createStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
model.feedAudioContent(stream_context,
np.frombuffer(frame, np.int16))
if ARGS.savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS.savewav:
vad_audio.write_wav(
os.path.join(
ARGS.savewav,
datetime.now().strftime(
"savewav_%Y-%m-%d_%H-%M-%S_%f.wav")), wav_data)
wav_data = bytearray()
text = model.finishStream(stream_context)
print("Recognized: %s" % text)
stream_context = model.createStream()
# publish
rospy.loginfo(text)
pub.publish(text)
def __init__(self, model, scorer=None):
logging.basicConfig(level=20)
print('Initializing model...')
logging.info("Model: %s", model)
self.model = deepspeech.Model(model)
if scorer:
logging.info("Scorer: %s", scorer)
self.model.enableExternalScorer(scorer)
def init_stt(output_graph_path, scorer_path):
# global model? How are we supposed to do multiprocessing then?
global model
model = deepspeech.Model(output_graph_path)
# It definitely seems very reasonable to adapt the
# ctc_decoder_with_kenlm tensorflow operator to use this...
model.enableExternalScorer(scorer_path)
logging.debug('Process {}: Loaded models'.format(os.getpid()))
def onStart(self):
super().onStart()
if not self.checkLanguage():
self.downloadLanguage()
self._model = deepspeech.Model(f'{self._langPath}/deepspeech-0.6.1-models/output_graph.tflite', 500)
self._model.enableDecoderWithLM(f'{self._langPath}/deepspeech-0.6.1-models/lm.binary', f'{self._langPath}/deepspeech-0.6.1-models/trie', 0.75, 1.85)
def load_deepspeech_model(self):
N_FEATURES = 25
N_CONTEXT = 9
BEAM_WIDTH = 500
LM_ALPHA = 0.75
LM_BETA = 1.85
ds = deepspeech.Model('deepspeech_model/deepspeech-0.7.3-models.pbmm')
return ds
def main(ARGS):
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.alphabet = os.path.join(
model_dir, ARGS.alphabet if ARGS.alphabet else 'alphabet.txt')
ARGS.lm = os.path.join(model_dir, ARGS.lm)
ARGS.trie = os.path.join(model_dir, ARGS.trie)
#time.sleep(30)
to_node("status", "Initializing model...")
# print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
logging.info("ARGS.alphabet: %s", ARGS.alphabet)
model = deepspeech.Model(ARGS.model, ARGS.n_features, ARGS.n_context,
ARGS.alphabet, ARGS.beam_width)
if ARGS.lm and ARGS.trie:
logging.info("ARGS.lm: %s", ARGS.lm)
logging.info("ARGS.trie: %s", ARGS.trie)
model.enableDecoderWithLM(ARGS.alphabet, ARGS.lm, ARGS.trie,
ARGS.lm_alpha, ARGS.lm_beta)
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate)
to_node("status", "Listening")
#print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not ARGS.nospinner: spinner = Halo(spinner='line')
stream_context = model.setupStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
model.feedAudioContent(stream_context,
np.frombuffer(frame, np.int16))
if ARGS.savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS.savewav:
vad_audio.write_wav(
os.path.join(
ARGS.savewav,
datetime.now().strftime(
"savewav_%Y-%m-%d_%H-%M-%S_%f.wav")), wav_data)
wav_data = bytearray()
text = model.finishStream(stream_context)
to_node("result", "{}".format(text))
# print("Recognized: %s" % text)
stream_context = model.setupStream()
def main(ARGS):
#pdb.set_trace()
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.scorer = os.path.join(model_dir, ARGS.scorer)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
model = deepspeech.Model(ARGS.model)
if ARGS.scorer:
logging.info("ARGS.scorer: %s", ARGS.scorer)
model.enableExternalScorer(ARGS.scorer)
audio_file = ARGS.file[:-4] + '.wav'
command = "ffmpeg -i {} -ab 160k -ac 1 -ar 16000 -vn {}".format(
ARGS.file, audio_file)
subprocess.call(command, shell=True)
#audio_file =
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate,
file=audio_file)
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not ARGS.nospinner:
spinner = Halo(spinner='line')
stream_context = model.createStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
stream_context.feedAudioContent(np.frombuffer(frame, np.int16))
if ARGS.savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS.savewav:
vad_audio.write_wav(
os.path.join(
ARGS.savewav,
datetime.now().strftime(
"savewav_%Y-%m-%d_%H-%M-%S_%f.wav")), wav_data)
wav_data = bytearray()
text = stream_context.finishStream()
print("Recognized: %s" % text)
if ARGS.keyboard:
from pyautogui import typewrite
typewrite(text)
stream_context = model.createStream()
def __init__(self, model_file, scorer_file, vad_audio):
log.info("DeepSpeech model: {}".format(model_file))
self.model = deepspeech.Model(model_file)
log.info("DeepSpeech scorer: {}".format(scorer_file))
self.model.enableExternalScorer(scorer_file)
self.spinner = Halo(spinner='line')
self.vad_audio = vad_audio
self.frames = vad_audio.vad_collector()
def TRANSCRIBING_SERVICE(): if transcribing_service._instance is None: transcribing_service._instance=transcribing_service() transcribing_service.model=deepspeech.Model(model_file_path) transcribing_service.model.enableExternalScorer(scorer_file_path) transcribing_service.model.setScorerAlphaBeta(lm_alpha, lm_beta) transcribing_service.model.setBeamWidth(beam_width) return transcribing_service._instance
def main(ARGS):
# Load DeepSpeech model
if os.path.isdir(ARGS.model):
model_dir = ARGS.model
ARGS.model = os.path.join(model_dir, 'output_graph.pb')
ARGS.scorer = os.path.join(model_dir, ARGS.scorer)
print('Initializing model...')
logging.info("ARGS.model: %s", ARGS.model)
model = deepspeech.Model(ARGS.model)
if ARGS.scorer:
logging.info("ARGS.scorer: %s", ARGS.scorer)
model.enableExternalScorer(ARGS.scorer)
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness,
device=ARGS.device,
input_rate=ARGS.rate,
file=ARGS.file)
print("Listening (ctrl-C to exit)...")
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not ARGS.nospinner:
spinner = Halo(spinner='line')
stream_context = model.createStream()
wav_data = bytearray()
for frame in frames:
if frame is not None:
if spinner: spinner.start()
logging.debug("streaming frame")
stream_context.feedAudioContent(np.frombuffer(frame, np.int16))
if ARGS.savewav: wav_data.extend(frame)
else:
if spinner: spinner.stop()
logging.debug("end utterence")
if ARGS.savewav:
vad_audio.write_wav("output.wav", wav_data)
wav_data = bytearray()
audio_file = open("output.wav", "rb")
response = speech_to_text.recognize(
audio=audio_file,
content_type='audio/wav',
timestamps=True,
word_confidence=True,
smart_formatting=True).get_result()
print(response['results'])
text_output = response['results'][0]['alternatives'][0][
'transcript']
text = text_output.strip()
#text = stream_context.finishStream()
print("Recognized: %s" % text)
stream_context = model.createStream()
def main(args):
# Initialize Ros Node and the Topic Publisher
rospy.init_node(ROS_NODE_NAME)
publisher = rospy.Publisher(ROS_PUBLISHER_TOPIC_NAME,
String,
queue_size=10)
# Load DeepSpeech model
if os.path.isdir(args.model):
model_dir = args.model
args.model = os.path.join(model_dir, 'output_graph.pb')
args.scorer = os.path.join(model_dir, args.scorer)
model = deepspeech.Model(args.model)
if args.scorer:
model.enableExternalScorer(args.scorer)
# Add hot words, boost level can be (-inf, +inf)
file = open(HOTWORDS_FILEPATH, "r")
lines = file.readlines()
for line in lines:
hot_word, boost_value = line.split(",")
model.addHotWord(hot_word, float(boost_value))
# Start audio with VAD
vad_audio = VADAudio(aggressiveness=args.vad_aggressiveness,
device=args.device,
input_rate=args.rate,
file=None)
print("ROS node '%s' started. Listening for speech (ctrl-C to exit)..." %
ROS_NODE_NAME)
frames = vad_audio.vad_collector()
# Stream from microphone to DeepSpeech using VAD
spinner = None
if not args.nospinner:
spinner = Halo(spinner='line')
stream_context = model.createStream()
for frame in frames:
if not rospy.is_shutdown():
if frame is not None:
if spinner: spinner.start()
stream_context.feedAudioContent(np.frombuffer(frame, np.int16))
else:
if spinner: spinner.stop()
recognized_text = stream_context.finishStream()
if recognized_text:
recognized_text = clean_text(recognized_text)
print("Recognized: %s" % recognized_text)
publisher.publish(recognized_text)
stream_context = model.createStream()
else:
stream_context.freeStream()
print("Ctrl-C received. Shutting down ROS node '%s'!" %
ROS_NODE_NAME)
break
def __init__(self, model=None, scorer=None):
self.sessions = dict()
#hold transcription sessions and the deepspeech model class statefully here, then the endpoint is stateless
self.sessions["last_used_id"] = -1
self.wait_time = 60 #amount of seconds to wait before killing a stream
self.timeout_check_time = 20 #number seconds to check if a stream has gone over its timout time since last action
self.timeout() #continues to repeat indefinitly
if model is None:
path = pathlib.Path(__file__).parent.absolute()
model_path = os.path.join(path, "./deepspeech-0.8.2-models.pbmm")
self.model = deepspeech.Model(model_path)
else:
self.model = deepspeech.Model(model)
if scorer is None:
path = pathlib.Path(__file__).parent.absolute()
scorer_path = os.path.join(path, "./deepspeech-0.8.2-models.scorer")
self.model.enableExternalScorer(scorer_path)
else:
self.model.enableExternalScorer(scorer)
