def create_asr():
import config
from kaldi.utils import lattice_to_nbest, wst2dict
from kaldi.decoders import PyOnlineLatgenRecogniser
from asr_utils import lattice_calibration
recogniser = PyOnlineLatgenRecogniser()
recogniser.setup(config.kaldi_config)
dictionary = wst2dict(config.wst_path)
path_to_text = PathToText(dictionary)
to_nbest = ToNBest(path_to_text, lattice_to_nbest, lattice_calibration)
to_best_path = ToBestPath(path_to_text)
return ASR(recogniser, to_nbest, to_best_path)
def decode_wrap(argv, audio_batch_size, wav_paths, file_output, wst_path=None):
wst = wst2dict(wst_path)
d = PyOnlineLatgenRecogniser()
d.setup(argv)
for wav_name, wav_path in wav_paths:
sw, sr = 2, 16000 # 16-bit audio so 1 sample_width = 2 chars
pcm = load_wav(wav_path, def_sample_width=sw, def_sample_rate=sr)
print('%s has %f sec' % (wav_name, (float(len(pcm)) / sw) / sr))
lat, lik, decoded_frames = decode(d, pcm)
lat.isyms = lat.osyms = fst.read_symbols_text(wst_path)
if DEBUG:
with open('pykaldi_%s.svg' % wav_name, 'w') as f:
f.write(lat._repr_svg_())
lat.write('%s_pykaldi.fst' % wav_name)
print(
"Log-likelihood per frame for utterance %s is %f over %d frames" %
(wav_name, int(lik / decoded_frames), decoded_frames))
word_ids = lattice_to_nbest(lat, n=10)
write_decoded(file_output, wav_name, word_ids, wst)
def decode_wrap(argv, audio_batch_size, wav_paths,
file_output, wst_path=None):
wst = wst2dict(wst_path)
d = PyOnlineLatgenRecogniser()
d.setup(argv)
for wav_name, wav_path in wav_paths:
sw, sr = 2, 16000 # 16-bit audio so 1 sample_width = 2 chars
pcm = load_wav(wav_path, def_sample_width=sw, def_sample_rate=sr)
print '%s has %f sec' % (wav_name, (float(len(pcm)) / sw) / sr)
lat, lik, decoded_frames = decode(d, pcm)
lat.isyms = lat.osyms = fst.read_symbols_text(wst_path)
if DEBUG:
with open('pykaldi_%s.svg' % wav_name, 'w') as f:
f.write(lat._repr_svg_())
lat.write('%s_pykaldi.fst' % wav_name)
print "Log-likelihood per frame for utterance %s is %f over %d frames" % (
wav_name, (lik / decoded_frames), decoded_frames)
word_ids = lattice_to_nbest(lat, n=10)
write_decoded(file_output, wav_name, word_ids, wst)
decoded = ' '.join([wst[w] for w in best_path])
else:
decoded = 'Empty hypothesis'
print(
"%s secs, frames: %d, prob: %f, %s " %
(str(time.time() - start), self.utt_frames, prob, decoded))
self.utt_frames = 0
self.d.reset(keep_buffer_data=False)
if self.dialog_end:
self.save_wav()
break
def save_wav(self):
wf = wave.open('live-demo-record.wav', 'wb')
wf.setnchannels(CHANNELS)
wf.setframerate(RATE)
wf.setsampwidth(self.pin.get_sample_size(FORMAT))
wf.writeframes(b''.join(self.frames))
wf.close()
if __name__ == '__main__':
audio_batch_size, wst_path = int(sys.argv[1]), sys.argv[2]
argv = sys.argv[3:]
print('Python args: %s' % str(sys.argv), file=sys.stderr)
wst = wst2dict(wst_path)
demo = LiveDemo(audio_batch_size, wst, argv)
demo.setup()
demo.run()
def create_dictionary(basedir):
global wst
wst = wst2dict('%s/models/words.txt' % basedir)
d.reset(keep_buffer_data=False)
return result
def get_audio_callback():
"""Returns a callback - function which handle incomming audio"""
def frame_in(in_data, frame_count, time_info, status):
d.frame_in(in_data)
return in_data, pyaudio.paContinue
return frame_in
@app.route('/')
def index():
return render_template('index.html')
audio_batch_size, wst_path = int(sys.argv[1]), sys.argv[2]
argv = sys.argv[3:]
print >> sys.stderr, 'Python args: %s' % str(sys.argv)
wst = wst2dict(wst_path)
d.setup(argv)
pin = pyaudio.PyAudio()
stream = pin.open(format=FORMAT, channels=CHANNELS,
rate=RATE, input=True, frames_per_buffer=audio_batch_size,
stream_callback=get_audio_callback())
app.run(host='0.0.0.0', debug=True)
