def test_asr_kaldi(self): asr = ASR(engine=ASR_ENGINE_NNET3) wavf = wave.open(TEST_WAVE_EN, 'rb') # check format self.assertEqual(wavf.getnchannels(), 1) self.assertEqual(wavf.getsampwidth(), 2) # process file in 250ms chunks chunk_frames = 250 * wavf.getframerate() / 1000 tot_frames = wavf.getnframes() num_frames = 0 while num_frames < tot_frames: finalize = False if (num_frames + chunk_frames) < tot_frames: nframes = chunk_frames else: nframes = tot_frames - num_frames finalize = True frames = wavf.readframes(nframes) num_frames += nframes samples = struct.unpack_from('<%dh' % nframes, frames) s, l = asr.decode(samples, finalize, wavf.getframerate()) wavf.close() self.assertEqual(s.strip(), TEST_WAVE_EN_TS)
def __init__(self, source=None, volume=None, aggressiveness=None, model_dir=None, lang=None, config=CONFIG): EventEmitter.__init__(self) self.config = config # ensure default values for k in CONFIG["listener"]: if k not in self.config["listener"]: self.config["listener"][k] = CONFIG["listener"][k] volume = volume or self.config["listener"]["default_volume"] aggressiveness = aggressiveness or self.config["listener"][ "default_aggressiveness"] model_dir = model_dir or self.config["listener"]["default_model_dir"] self.lang = lang or self.config["lang"] if "-" in self.lang: self.lang = self.lang.split("-")[0] if "{lang}" in model_dir: model_dir = model_dir.format(lang=self.lang) if not isdir(model_dir): if model_dir in self._default_models: logging.error( "you need to install the package: " "kaldi-chain-zamia-speech-{lang}".format(lang=self.lang)) raise ModelNotFound self.rec = PulseRecorder(source_name=source, volume=volume) self.vad = VAD(aggressiveness=aggressiveness) logging.info("Loading model from %s ..." % model_dir) self.asr = ASR(engine=ASR_ENGINE_NNET3, model_dir=model_dir, kaldi_beam=self.config["listener"]["default_beam"], kaldi_acoustic_scale=self.config["listener"] ["default_acoustic_scale"], kaldi_frame_subsampling_factor=self.config["listener"] ["default_frame_subsampling_factor"]) self._hotwords = dict(self.config["hotwords"])
def test_asr_pocketsphinx(self): asr = ASR(engine=ASR_ENGINE_POCKETSPHINX, model_dir=POCKETSPHINX_MODELDIR, model_name=POCKETSPHINX_MODELNAME) wavf = wave.open(TEST_WAVE_EN, 'rb') # check format self.assertEqual(wavf.getnchannels(), 1) self.assertEqual(wavf.getsampwidth(), 2) # process file in 250ms chunks chunk_frames = 250 * wavf.getframerate() / 1000 tot_frames = wavf.getnframes() num_frames = 0 while num_frames < tot_frames: finalize = False if (num_frames + chunk_frames) < tot_frames: nframes = chunk_frames else: nframes = tot_frames - num_frames finalize = True frames = wavf.readframes(nframes) num_frames += nframes samples = struct.unpack_from('<%dh' % nframes, frames) s, l = asr.decode(wavf.getframerate(), samples, finalize) if not finalize: self.assertEqual(s, None) wavf.close() self.assertEqual(s.strip(), TEST_WAVE_EN_TS_PS)
class KaldiWWSpotter(EventEmitter): _default_models = ["/opt/kaldi/model/kaldi-generic-en-tdnn_250", "/opt/kaldi/model/kaldi-generic-de-tdnn_250"] def __init__(self, source=None, volume=None, aggressiveness=None, model_dir=None, lang=None, config=CONFIG): EventEmitter.__init__(self) self.config = config # ensure default values for k in CONFIG["listener"]: if k not in self.config["listener"]: self.config["listener"][k] = CONFIG["listener"][k] volume = volume or self.config["listener"]["default_volume"] aggressiveness = aggressiveness or self.config["listener"][ "default_aggressiveness"] model_dir = model_dir or self.config["listener"]["default_model_dir"] self.lang = lang or self.config["lang"] if "-" in self.lang: self.lang = self.lang.split("-")[0] if "{lang}" in model_dir: model_dir = model_dir.format(lang=self.lang) if not isfile(model_dir): if model_dir in self._default_models: logging.error("you need to install the package: " "kaldi-chain-zamia-speech-{lang}".format( lang=self.lang)) raise ModelNotFound self.rec = PulseRecorder(source_name=source, volume=volume) self.vad = VAD(aggressiveness=aggressiveness) logging.info("Loading model from %s ..." % model_dir) self.asr = ASR(engine=ASR_ENGINE_NNET3, model_dir=model_dir, kaldi_beam=self.config["listener"]["default_beam"], kaldi_acoustic_scale=self.config["listener"][ "default_acoustic_scale"], kaldi_frame_subsampling_factor=self.config["listener"][ "default_frame_subsampling_factor"]) self._hotwords = dict(self.config["hotwords"]) def add_hotword(self, name, config=None): config = config or {"transcriptions": [name], "intent": name} self._hotwords[name] = config def remove_hotword(self, name): if name in self._hotwords.keys(): self._hotwords.pop(name) @property def hotwords(self): return self._hotwords def _detection_event(self, message_type, message_data): serialized_message = json.dumps( {"type": message_type, "data": message_data}) logging.debug(serialized_message) self.emit(message_type, serialized_message) def _process_transcription(self, user_utt, confidence=0.99): for hotw in self.hotwords: if not self.hotwords[hotw].get("active"): continue rule = self.hotwords[hotw].get("rule", "sensitivity") s = 1 - self.hotwords[hotw].get("sensitivity", 0.2) confidence = (confidence + s) / 2 for w in self.hotwords[hotw]["transcriptions"]: if (w in user_utt and rule == "in") or \ (user_utt.startswith(w) and rule == "start") or \ (user_utt.endswith(w) and rule == "end") or \ (fuzzy_match(w, user_utt) >= s and rule == "sensitivity") or \ (w == user_utt and rule == "equal"): yield {"hotword": hotw, "utterance": user_utt, "confidence": confidence, "intent": self.hotwords[hotw]["intent"]} def _detect_ww(self, user_utt, confidence=0.99): for hw_data in self._process_transcription(user_utt, confidence): sound = self.hotwords[hw_data["hotword"]].get("sound") if sound and isfile(sound): play_sound(sound) self._detection_event("hotword", hw_data) def decode_wav_file(self, wav_file): user_utt, confidence = self.asr.decode_wav_file(wav_file) confidence = 1 - exp(-1 * confidence) return user_utt, confidence def wav_file_hotwords(self, wav_file): user_utt, confidence = self.decode_wav_file(wav_file) return list(self._process_transcription(user_utt, confidence)) def run(self): self.rec.start_recording() logging.info("Listening") while True: samples = self.rec.get_samples() audio, finalize = self.vad.process_audio(samples) if not audio: continue logging.debug('decoding audio len=%d finalize=%s audio=%s' % ( len(audio), repr(finalize), audio[0].__class__)) user_utt, confidence = self.asr.decode(audio, finalize, stream_id="mic") confidence = 1 - exp(-1 * confidence) if finalize and user_utt: self._detection_event("transcription", {"utterance": user_utt, "confidence": confidence}) self._detect_ww(user_utt, confidence)
rec = PulseRecorder (source_name=source, volume=volume) # # VAD # vad = VAD(aggressiveness=aggressiveness) # # ASR # print "Loading model from %s ..." % model_dir asr = ASR(engine = ASR_ENGINE_NNET3, model_dir = model_dir, kaldi_beam = DEFAULT_BEAM, kaldi_acoustic_scale = DEFAULT_ACOUSTIC_SCALE, kaldi_frame_subsampling_factor = DEFAULT_FRAME_SUBSAMPLING_FACTOR) # # main # rec.start_recording() print "Please speak." while True: samples = rec.get_samples()
MODELDIR = '/opt/kaldi/model/kaldi-generic-en-tdnn_250' VOLUME = 150 class Intent(Enum): HELLO = 1 LIGHT = 2 RADIO = 3 print("Initializing...") radio_on = False lights_on = False asr = ASR(model_dir=MODELDIR) rec = PulseRecorder(volume=VOLUME) vad = VAD() tts = TTS(engine="espeak", voice="en") utt_map = {} def add_utt(utterance, intent): utt_map[utterance] = intent add_utt("hello computer", Intent.HELLO) add_utt("switch on the lights", Intent.LIGHT) add_utt("switch off the lights", Intent.LIGHT) add_utt("switch on the radio", Intent.RADIO)
# kernal.setup_align_utterances(lang=lang) paint_main() logging.debug ('AI kernal initialized.') # # context # cur_context = kernal.find_prev_context(USER_URI) # # ASR # misc.message_popup(stdscr, 'Initializing...', 'Init ASR...') asr = ASR(engine = ASR_ENGINE_NNET3, model_dir = kaldi_model_dir, model_name = kaldi_model) paint_main() logging.debug ('ASR initialized.') # # main loop # while True: paint_main() c = stdscr.getch() if c == ord('q'): break elif c == ord('r'):
# # setup AI DB, Kernal and Context # kernal = AIKernal.from_ini_file() for skill in kernal.all_skills: kernal.consult_skill(skill) kernal.setup_nlp_model() ctx = kernal.create_context() logging.debug('AI kernal initialized.') # # ASR # asr = ASR(model_dir=options.asr_model) logging.debug('ASR initialized.') # # TTS # tts = TTS(engine="espeak", voice="en") # # main loop # print(chr(27) + "[2J") while True:
lang = kernal.nlp_model.lang ctx = AIContext(USER_URI, kernal.session, lang, DEMO_REALM, kernal, test_mode=False) logging.debug('AI kernal initialized.') # # ASR # asr = ASR(engine=ASR_ENGINE_NNET3, model_dir=kaldi_model_dir, model_name=kaldi_model, kaldi_beam=kaldi_beam, kaldi_acoustic_scale=kaldi_acoustic_scale, kaldi_frame_subsampling_factor=kaldi_frame_subsampling_factor) logging.debug('ASR initialized.') # # TTS # tts = TTS(host_tts=tts_host, port_tts=tts_port, locale=tts_locale, voice=tts_voice, engine=tts_engine, speed=tts_speed, pitch=tts_pitch)
logging.debug('AI kernal initialized.') # # context # cur_context = kernal.find_prev_context(USER_URI) # # ASR # misc.message_popup(stdscr, 'Initializing...', 'Init ASR...') asr = ASR(engine=ASR_ENGINE_NNET3, model_dir=kaldi_model_dir, model_name=kaldi_model, kaldi_beam=kaldi_beam, kaldi_acoustic_scale=kaldi_acoustic_scale, kaldi_frame_subsampling_factor=kaldi_frame_subsampling_factor) paint_main() logging.debug('ASR initialized.') # # main loop # while True: paint_main() c = stdscr.getch() if c == ord('q'):
#!/usr/bin/env python3 from nltools.asr import ASR MODELDIR = '/opt/kaldi/model/kaldi-generic-en-tdnn_250' WAVFILE = 'dw961.wav' asr = ASR(model_dir=MODELDIR) s, l = asr.decode_wav_file(WAVFILE) print("Decoded %s: %s" % (WAVFILE, s))
def test_asr_kaldi_wavefile(self): asr = ASR(engine=ASR_ENGINE_NNET3) s, l = asr.decode_wav_file(TEST_WAVE_EN) self.assertEqual(s.strip(), TEST_WAVE_EN_TS)
def test_asr_pocketsphinx_wavefile(self): asr = ASR(engine=ASR_ENGINE_POCKETSPHINX, model_dir=POCKETSPHINX_MODELDIR, model_name=POCKETSPHINX_MODELNAME) s, l = asr.decode_wav_file(TEST_WAVE_EN) self.assertEqual(s.strip(), TEST_WAVE_EN_TS_PS)
# # setup AI DB, Kernal and Context # kernal = AIKernal.from_ini_file() for skill in kernal.all_skills: kernal.consult_skill (skill) kernal.setup_nlp_model() ctx = kernal.create_context() logging.debug ('AI kernal initialized.') # # ASR # asr = ASR(model_dir = options.asr_model) logging.debug ('ASR initialized.') # # TTS # tts = TTS(engine="espeak", voice="en") # # main loop # print(chr(27) + "[2J") while True: