def re_run(self, utt):
if 'wavpath' not in utt:
return
k = Kaldi(
get_resource('data/nnet_a_gpu_online'),
self.gen_hclg_filename,
get_resource('PROTO_LANGDIR'))
audio = numm3.sound2np(
os.path.join(self.resources['attach'].attachdir, utt['wavpath']),
nchannels=1,
R=8000)
k.push_chunk(audio.tostring())
wds = k.get_final()
k.stop()
for wd in wds:
del wd['phones']
utt['command_words'] = wds
utt['command'] = ' '.join([X['word'] for X in wds])
reactor.callFromThread(self.db.onchange, None, {"type": "change",
"id": utt["_id"],
"doc": utt})
def re_run(self, utt):
if 'wavpath' not in utt:
return
k = Kaldi(get_resource('data/nnet_a_gpu_online'),
self.gen_hclg_filename, get_resource('PROTO_LANGDIR'))
audio = numm3.sound2np(os.path.join(self.resources['attach'].attachdir,
utt['wavpath']),
nchannels=1,
R=8000)
k.push_chunk(audio.tostring())
wds = k.get_final()
k.stop()
for wd in wds:
del wd['phones']
utt['command_words'] = wds
utt['command'] = ' '.join([X['word'] for X in wds])
reactor.callFromThread(self.db.onchange, None, {
"type": "change",
"id": utt["_id"],
"doc": utt
})
self._cmd("reset")
def stop(self):
if not self.finished:
self.finished = True
self._cmd("stop")
self._p.stdin.close()
self._p.stdout.close()
self._p.wait()
def __del__(self):
self.stop()
if __name__ == '__main__':
import numm3
import sys
infile = sys.argv[1]
k = Kaldi()
buf = numm3.sound2np(infile, nchannels=1, R=8000)
print('loaded_buf', len(buf))
idx = 0
while idx < len(buf):
k.push_chunk(buf[idx:idx + 160000].tostring())
print(k.get_final())
idx += 160000
def _reset(self):
self._cmd("reset")
def stop(self):
if not self.finished:
self.finished = True
self._cmd("stop")
self._p.stdin.close()
self._p.stdout.close()
self._p.wait()
def __del__(self):
self.stop()
if __name__=='__main__':
import numm3
import sys
infile = sys.argv[1]
k = Kaldi()
buf = numm3.sound2np(infile, nchannels=1, R=8000)
print('loaded_buf', len(buf))
idx=0
while idx < len(buf):
k.push_chunk(buf[idx:idx+160000].tostring())
print(k.get_final())
idx += 160000
hit_results[utt_idx] = hit_results.get(utt_idx, 0) + 1
else:
self.onpreview(ret, utt_idx)
def stop(self):
self.stopped = True
self.utts[-1].stop()
def join(self):
self.t.join()
if __name__=='__main__':
import sys
# Simulate with an audio file
AUDIOFILE = sys.argv[1]
OUTDIR = sys.argv[2]
sess = Session(OUTDIR)
test_audio = numm3.sound2np(AUDIOFILE, nchannels=1, R=8000)
cur_start = 0
BUF_LEN = 200
while True:
sess.feed(test_audio[cur_start:cur_start+BUF_LEN])
cur_start += BUF_LEN
if cur_start >= len(test_audio):
break
sess.stop()
sess.join()
import math
import sys
AUDIOPATH = sys.argv[1]
JSON_OUT = sys.argv[2]
N_THREADS = 4
T_PER_CHUNK = 10
OVERLAP_T = 2
kaldi_queue = Queue()
for i in range(N_THREADS):
kaldi_queue.put(standard_kaldi.Kaldi())
# Preload entire audio
audiobuf = numm3.sound2np(AUDIOPATH, R=8000, nchannels=1)
n_chunks = int(math.ceil(len(audiobuf) / (8000.0 * (T_PER_CHUNK-OVERLAP_T))))
print 'sharding into %d chunks' % (n_chunks)
chunks = [] # (idx, [words])
def transcribe_chunk(idx):
st = idx * (T_PER_CHUNK-OVERLAP_T) * 8000
end= st + T_PER_CHUNK * 8000
buf = audiobuf[st:end]
print buf.shape
k = kaldi_queue.get()