def test_to_csv():
tests = [
[{"words": []}, '', 'empty'],
[
{
"words": [
{"case": "success", "word": "A", "alignedWord": "a", "start": 0, "end": 1},
]
},
'A,a,0,1\r\n',
'single word',
],
[
{
"words": [
{"case": "success", "word": "A", "alignedWord": "a", "start": 0, "end": 1},
{"case": "not-found-in-audio", "word": "B", "alignedWord": "b", "start": 2, "end": 3},
]
},
'A,a,0,1\r\nB,b,2,3\r\n',
'multi-word',
],
[
{
"words": [
{"case": "not-found-in-audio", "word": "A", "alignedWord": "a", "start": 0, "end": 1},
]
},
'A,a,0,1\r\n',
'not found in audio',
],
[
{
"words": [
{"case": "not-found-in-transcript", "word": "A", "alignedWord": "a", "start": 0, "end": 1},
]
},
'',
'not found in transcript',
],
]
for test in tests:
input, want, name = test
got = to_csv(input)
assert_equals(want, got)
def save():
with open(os.path.join(outdir, 'align.json'), 'w') as jsfile:
json.dump(output, jsfile, indent=2)
with open(os.path.join(outdir, 'align.csv'), 'w') as csvfile:
csvfile.write(to_csv(output))
def transcribe(self, uid, transcript, audio, async):
proto_langdir = get_resource('PROTO_LANGDIR')
status = self.get_status(uid)
status['status'] = 'STARTED'
output = {
'transcript': transcript
}
outdir = os.path.join(self.data_dir, 'transcriptions', uid)
tran_path = os.path.join(outdir, 'transcript.txt')
with codecs.open(tran_path, 'w', 'utf-8') as tranfile:
tranfile.write(transcript)
audio_path = os.path.join(outdir, 'upload')
with open(audio_path, 'w') as wavfile:
wavfile.write(audio)
status['status'] = 'ENCODING'
wavfile = os.path.join(outdir, 'a.wav')
if to_wav(os.path.join(outdir, 'upload'), wavfile) != 0:
status['status'] = 'ERROR'
status['error'] = "Encoding failed. Make sure that you've uploaded a valid media file."
# Save the status so that errors are recovered on restart of the server
# XXX: This won't work, because the endpoint will override this file
with open(os.path.join(outdir, 'status.json'), 'w') as jsfile:
json.dump(status, jsfile, indent=2)
return
#XXX: Maybe we should pass this wave object instead of the
# file path to align_progress
wav_obj = wave.open(wavfile, 'r')
status['duration'] = wav_obj.getnframes() / float(wav_obj.getframerate())
status['status'] = 'TRANSCRIBING'
def on_progress(p):
for k,v in p.items():
status[k] = v
if len(transcript.strip()) > 0:
ms = metasentence.MetaSentence(transcript, self.vocab)
ks = ms.get_kaldi_sequence()
gen_hclg_filename = language_model.make_bigram_language_model(ks, proto_langdir)
kaldi_queue = Queue()
for i in range(self.nthreads):
kaldi_queue.put(standard_kaldi.Kaldi(
get_resource('data/nnet_a_gpu_online'),
gen_hclg_filename,
proto_langdir)
)
mtt = MultiThreadedTranscriber(kaldi_queue, nthreads=self.nthreads)
elif hasattr(self, 'full_transcriber'):
mtt = self.full_transcriber
else:
status['status'] = 'ERROR'
status['error'] = 'No transcript provided and no language model for full transcription'
return
words = mtt.transcribe(wavfile, progress_cb=on_progress)
output = {}
if len(transcript.strip()) > 0:
# Clear queue (would this be gc'ed?)
for i in range(self.nthreads):
k = kaldi_queue.get()
k.stop()
# Align words
output['words'] = diff_align.align(words, ms)
output['transcript'] = transcript
# Perform a second-pass with unaligned words
logging.info("%d unaligned words (of %d)" % (len([X for X in output['words'] if X.get("case") == "not-found-in-audio"]), len(output['words'])))
status['status'] = 'ALIGNING'
output['words'] = multipass.realign(wavfile, output['words'], ms, nthreads=self.nthreads, progress_cb=on_progress)
logging.info("after 2nd pass: %d unaligned words (of %d)" % (len([X for X in output['words'] if X.get("case") == "not-found-in-audio"]), len(output['words'])))
else:
# Match format
output = make_transcription_alignment({"words": words})
# ...remove the original upload
os.unlink(os.path.join(outdir, 'upload'))
# Save
with open(os.path.join(outdir, 'align.json'), 'w') as jsfile:
json.dump(output, jsfile, indent=2)
with open(os.path.join(outdir, 'align.csv'), 'w') as csvfile:
csvfile.write(to_csv(output))
# Inline the alignment into the index.html file.
htmltxt = open(get_resource('www/view_alignment.html')).read()
htmltxt = htmltxt.replace("var INLINE_JSON;", "var INLINE_JSON=%s;" % (json.dumps(output)));
open(os.path.join(outdir, 'index.html'), 'w').write(htmltxt)
status['status'] = 'OK'
logging.info('done with transcription.')
return output
def save():
with open(os.path.join(outdir, 'align.json'), 'w') as jsfile:
json.dump(output, jsfile, indent=2)
with open(os.path.join(outdir, 'align.csv'), 'w') as csvfile:
csvfile.write(to_csv(output))
def test_to_csv():
tests = [
[{
"words": []
}, '', 'empty'],
[
{
"words": [
{
"case": "success",
"word": "A",
"alignedWord": "a",
"start": 0,
"end": 1
},
]
},
'A,a,0,1\r\n',
'single word',
],
[
{
"words": [
{
"case": "success",
"word": "A",
"alignedWord": "a",
"start": 0,
"end": 1
},
{
"case": "not-found-in-audio",
"word": "B",
"alignedWord": "b",
"start": 2,
"end": 3
},
]
},
'A,a,0,1\r\nB,b,2,3\r\n',
'multi-word',
],
[
{
"words": [
{
"case": "not-found-in-audio",
"word": "A",
"alignedWord": "a",
"start": 0,
"end": 1
},
]
},
'A,a,0,1\r\n',
'not found in audio',
],
[
{
"words": [
{
"case": "not-found-in-transcript",
"word": "A",
"alignedWord": "a",
"start": 0,
"end": 1
},
]
},
'',
'not found in transcript',
],
]
for test in tests:
input, want, name = test
got = to_csv(input)
assert_equals(want, got)