class FullTranscriber():
def __init__(self, resources, nthreads=2):
self.available = False
if nthreads <= 0: return
if not os.path.exists(resources.full_hclg_path): return
queue = kaldi_queue.build(resources, nthreads=nthreads)
self.mtt = MultiThreadedTranscriber(queue, nthreads=nthreads)
self.available = True
def transcribe(self, wavfile, progress_cb=None, logging=None):
words, duration = self.mtt.transcribe(wavfile, progress_cb=progress_cb)
return self.make_transcription_alignment(words)
@staticmethod
def make_transcription_alignment(trans):
# Spoof the `diff_align` output format
transcript = ""
words = []
for t_wd in trans:
word = transcription.Word(
case=transcription.Word.SUCCESS,
startOffset=len(transcript),
endOffset=len(transcript) + len(t_wd.word),
word=t_wd.word,
alignedWord=t_wd.word,
phones=t_wd.phones,
start=t_wd.start,
end=t_wd.end)
words.append(word)
transcript += word.word + " "
return Transcription(words=words, transcript=transcript)
class FullTranscriber():
def __init__(self, resources, nthreads=2):
self.available = False
if nthreads <= 0: return
if not os.path.exists(resources.full_hclg_path): return
queue = kaldi_queue.build(resources, nthreads=nthreads)
self.mtt = MultiThreadedTranscriber(queue, nthreads=nthreads)
self.available = True
def transcribe(self, wavfile, progress_cb=None, logging=None):
words = self.mtt.transcribe(wavfile, progress_cb=progress_cb)
return self.make_transcription_alignment(words)
@staticmethod
def make_transcription_alignment(trans):
# Spoof the `diff_align` output format
transcript = ""
words = []
for t_wd in trans:
word = transcription.Word(case="success",
startOffset=len(transcript),
endOffset=len(transcript) +
len(t_wd.word),
word=t_wd.word,
alignedWord=t_wd.word,
phones=t_wd.phones,
start=t_wd.start,
end=t_wd.end)
words.append(word)
transcript += word.word + " "
return Transcription(words=words, transcript=transcript)
class ForcedAligner():
def __init__(self,
resources,
transcript,
nthreads=4,
context_width=3,
**kwargs):
self.kwargs = kwargs
self.nthreads = nthreads
self.transcript = transcript
self.resources = resources
self.context_width = context_width
self.ms = metasentence.MetaSentence(transcript, resources.vocab)
ks = self.ms.get_kaldi_sequence()
gen_hclg_filename = language_model.make_bigram_language_model(
ks, resources.proto_langdir, context_width=context_width, **kwargs)
self.queue = kaldi_queue.build(resources,
hclg_path=gen_hclg_filename,
nthreads=nthreads)
self.mtt = MultiThreadedTranscriber(self.queue, nthreads=nthreads)
def transcribe(self, wavfile, progress_cb=None, logging=None):
words, duration = self.mtt.transcribe(wavfile, progress_cb=progress_cb)
# Clear queue (would this be gc'ed?)
for i in range(self.nthreads):
k = self.queue.get()
k.stop()
# Align words
words = diff_align.align(words, self.ms, **self.kwargs)
# Perform a second-pass with unaligned words
if logging is not None:
logging.info(
"%d unaligned words (of %d)" %
(len([X
for X in words if X.not_found_in_audio()]), len(words)))
if progress_cb is not None:
progress_cb({'status': 'ALIGNING'})
words = multipass.realign(wavfile,
words,
self.ms,
context_width=self.context_width,
resources=self.resources,
nthreads=self.nthreads,
progress_cb=progress_cb)
if logging is not None:
logging.info(
"after 2nd pass: %d unaligned words (of %d)" %
(len([X
for X in words if X.not_found_in_audio()]), len(words)))
words = AdjacencyOptimizer(words, duration).optimize()
return Transcription(words=words, transcript=self.transcript)
def test_transcriber(self):
from gentle import resampled, kaldi_queue, standard_kaldi, Resources
from gentle.transcriber import MultiThreadedTranscriber
resources = Resources()
k_queue = kaldi_queue.build(resources, 1)
trans = MultiThreadedTranscriber(k_queue)
with resampled('examples/data/lucier.mp3', 10.5, 2.5) as filename:
words, duration = trans.transcribe(filename)
self.assertEqual(words[0].word, "different")
def test_transcriber(self):
import subprocess
from gentle import resampled, kaldi_queue, standard_kaldi, Resources
from gentle.transcriber import MultiThreadedTranscriber
standard_kaldi.STDERR = subprocess.STDOUT
resources = Resources()
k_queue = kaldi_queue.build(resources, 1)
trans = MultiThreadedTranscriber(k_queue)
with resampled(self.audio, 10.5, 2.5) as filename:
words, duration = trans.transcribe(filename)
self.assertEqual(words[0].word, "different")
def test_transcriber(self):
import subprocess
from gentle import resampled, kaldi_queue, standard_kaldi, Resources
from gentle.transcriber import MultiThreadedTranscriber
standard_kaldi.STDERR = subprocess.STDOUT
resources = Resources()
k_queue = kaldi_queue.build(resources, 1)
trans = MultiThreadedTranscriber(k_queue)
with resampled(self.audio, 10.5, 2.5) as filename:
words, duration = trans.transcribe(filename)
self.assertEqual(words[0].word, "different")
class ForcedAligner():
def __init__(self, resources, transcript, nthreads=4, **kwargs):
self.kwargs = kwargs
self.nthreads = nthreads
self.transcript = transcript
self.resources = resources
self.ms = metasentence.MetaSentence(transcript, resources.vocab)
ks = self.ms.get_kaldi_sequence()
gen_hclg_filename = language_model.make_bigram_language_model(ks, resources.proto_langdir, **kwargs)
self.queue = kaldi_queue.build(resources, hclg_path=gen_hclg_filename, nthreads=nthreads)
self.mtt = MultiThreadedTranscriber(self.queue, nthreads=nthreads)
def transcribe(self, wavfile, progress_cb=None, logging=None):
words, duration = self.mtt.transcribe(wavfile, progress_cb=progress_cb)
# Clear queue (would this be gc'ed?)
for i in range(self.nthreads):
k = self.queue.get()
k.stop()
# Align words
words = diff_align.align(words, self.ms, **self.kwargs)
# Perform a second-pass with unaligned words
if logging is not None:
logging.info("%d unaligned words (of %d)" % (len([X for X in words if X.not_found_in_audio()]), len(words)))
if progress_cb is not None:
progress_cb({'status': 'ALIGNING'})
words = multipass.realign(wavfile, words, self.ms, resources=self.resources, nthreads=self.nthreads, progress_cb=progress_cb)
if logging is not None:
logging.info("after 2nd pass: %d unaligned words (of %d)" % (len([X for X in words if X.not_found_in_audio()]), len(words)))
words = AdjacencyOptimizer(words, duration).optimize()
return Transcription(words=words, transcript=self.transcript)