def __init__(self, *args, **kwargs):
super(Tester, self).__init__(*args, **kwargs)
print('Initializing aspire model...')
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = 13
decoder_opts.max_active = 7000
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.acoustic_scale = 1.0
decodable_opts.frame_subsampling_factor = 3
decodable_opts.frames_per_chunk = 150
self.asr = NnetLatticeFasterRecognizer.from_files(
"/home/chris/git/pykaldi/examples/setups/aspire/exp/tdnn_7b_chain_online/final.mdl",
"/home/chris/git/pykaldi/examples/setups/aspire/exp/tdnn_7b_chain_online/graph_pp/HCLG.fst",
"/home/chris/git/pykaldi/examples/setups/aspire/data/lang/words.txt",
decoder_opts=decoder_opts,
decodable_opts=decodable_opts)
_, fn = tempfile.mkstemp()
os.remove(fn)
self.scp_fn = scp_fn = '%s.scp' % fn
# Define feature pipelines as Kaldi rspecifiers
self.feats_rspec = (
f"ark:compute-mfcc-feats --config=/home/chris/git/pykaldi/examples/setups/aspire/conf/mfcc_hires.conf scp:{scp_fn} ark:- |"
)
self.ivectors_rspec = (
f"ark:compute-mfcc-feats --config=/home/chris/git/pykaldi/examples/setups/aspire/conf/mfcc_hires.conf scp:{scp_fn} ark:- |"
f"ivector-extract-online2 --config=/home/chris/git/pykaldi/examples/setups/aspire/conf/ivector_extractor.conf " \
f"ark:/home/chris/git/pykaldi/examples/setups/aspire/data/test/spk2utt ark:- ark:- |"
)
def __init__(self, scp, model, graph, words, conf, iconf, spk2utt, output, printed=False, log=False):
"""
Инициализация транскриптора
Аргументы:
scp: путь к .SCP файлу с аудио
model: путь к .MDL файлу модели распознавания
graph: путь к .FST файлу общего графа распознавания
words: путь к .TXT файлу текстового корпуса
conf: путь к .CONF конфигурационному файлу распознавания
iconf: путь к .CONF конфигурационному файлу векторного экстрактора
spk2utt: путь к файлу перечисления сегментов для каждого говорящего
output: путь к директории с результатами распознавания
printed: признак печати результатов распознавания
log: признак логирования
"""
self.scp = scp
self.model = model
self.graph = graph
self.words = words
self.conf = conf
self.iconf = iconf
self.spk2utt = spk2utt
self.output = Path(output)
self.printed = printed
self.log = log
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = 13
decoder_opts.max_active = 7000
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.acoustic_scale = 1.0
decodable_opts.frame_subsampling_factor = 3
self.asr = NnetLatticeFasterRecognizer.from_files(self.model, self.graph, self.words,
decoder_opts=decoder_opts, decodable_opts=decodable_opts)
def __init__(self, scp, model, post, conf, output, log=False):
"""
Инициализация сегментатора
Аргументы:
scp: путь к .SCP файлу с аудио
model: путь к .RAW файлу модели сегментации
post: путь к .VEC файлу апостериорных вероятностей сегментации
conf: путь к .CONF конфигурационному файлу сегментации
output: путь к директории с результатами сегментации
log: признак логирования
"""
self.scp = scp
self.model = model
self.post = post
self.conf = conf
self.output = Path(output)
self.log = log
sad_model = NnetSAD.read_model(model)
sad_post = NnetSAD.read_average_posteriors(post)
sad_transform = NnetSAD.make_sad_transform(sad_post)
sad_graph = NnetSAD.make_sad_graph()
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.extra_left_context = 79
decodable_opts.extra_right_context = 21
decodable_opts.extra_left_context_initial = 0
decodable_opts.extra_right_context_final = 0
decodable_opts.frames_per_chunk = 150
decodable_opts.acoustic_scale = 0.3
self.sad = NnetSAD(sad_model,
sad_transform,
sad_graph,
decodable_opts=decodable_opts)
self.seg = SegmentationProcessor([2])
#!/usr/bin/env python
from __future__ import print_function
from kaldi.asr import NnetLatticeFasterRecognizer
from kaldi.decoder import LatticeFasterDecoderOptions
from kaldi.nnet3 import NnetSimpleComputationOptions
from kaldi.util.table import SequentialMatrixReader
# Construct recognizer
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = 13
decoder_opts.max_active = 7000
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.acoustic_scale = 1.0
decodable_opts.frame_subsampling_factor = 3
decodable_opts.frames_per_chunk = 150
asr = NnetLatticeFasterRecognizer.from_files(
"exp/tdnn_7b_chain_online/final.mdl",
"exp/tdnn_7b_chain_online/graph_pp/HCLG.fst",
"exp/tdnn_7b_chain_online/graph_pp/words.txt",
decoder_opts=decoder_opts,
decodable_opts=decodable_opts)
# Define feature pipelines as Kaldi rspecifiers
feats_rspec = (
"ark:compute-mfcc-feats --config=conf/mfcc.conf scp:data/wav.scp ark:- |")
ivectors_rspec = (
"ark:compute-mfcc-feats --config=conf/mfcc.conf scp:data/wav.scp ark:- |"
"ivector-extract-online2 --config=conf/ivector.conf ark:data/spk2utt ark:- ark:- |"
)
def main(self):
# Construct recognizer
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = 13
decoder_opts.max_active = 7000
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.acoustic_scale = 1.0
decodable_opts.frame_subsampling_factor = 3
decodable_opts.frames_per_chunk = 150
asr = NnetLatticeFasterRecognizer.from_files(
self.dir_path + "/exp/tdnn_7b_chain_online/final.mdl",
self.dir_path + "/new/graph/HCLG.fst",
self.dir_path + "/new/graph/words.txt",
decoder_opts=decoder_opts,
decodable_opts=decodable_opts)
p = pyaudio.PyAudio()
# ############################################
# sentiment_analyzer = SentimentAnalyzer(self.dir_path)
# model = load_model(self.dir_path + '/lstm.h5')
############################################
stream = p.open(format=FORMAT,
channels=CHANNELS,
rate=RATE,
input=True,
frames_per_buffer=CHUNK)
print("* recording")
audio2send = []
cur_data = '' # current chunk of audio data
rel = RATE / CHUNK
slid_win = deque(maxlen=int(SILENCE_LIMIT * rel) + 1)
# Prepend audio from 0.5 seconds before noise was detected
prev_audio = deque(maxlen=int(PREV_AUDIO * rel) + 1)
started = False
n = num_phrases
response = []
while num_phrases == -1 or n > 0:
cur_data = stream.read(CHUNK)
slid_win.append(math.sqrt(abs(audioop.avg(cur_data, 4))))
# print slid_win[-1]
if sum([x > THRESHOLD for x in slid_win]) > 0:
if (not started):
# print "Starting record of phrase"
started = True
audio2send.append(cur_data)
elif (started is True):
# print "Finished"
# The limit was reached, finish capture and deliver.
filename = self.save_speech(list(prev_audio) + audio2send, p, self.SAVE_PATH, self.WAVE_OUTPUT_FILENAME)
# Send file to Google and get response
r = self.recognize_speech(asr)
if num_phrases == -1:
print("Detected speech: ", r)
# if r != None:
# sentiment_analyzer.get_sentiment(r, model)
else:
response.append(r)
# Remove temp file. Comment line to review.
os.remove(filename)
# Reset all
started = False
slid_win = deque(maxlen=int(SILENCE_LIMIT * rel) + 1)
prev_audio = deque(maxlen=int(0.5 * rel) + 1)
audio2send = []
n -= 1
else:
prev_audio.append(cur_data)
stream.stop_stream()
stream.close()
p.terminate()
def init(self, nnet_directory, transcription_directory):
return_msg = "KaldiDecoder:init"
debug_data = []
feats = ""
ivectors = ""
decoder_opts = None
decodable_opts = None
asr = None
## input validation
if nnet_directory is not None:
if type(nnet_directory) is not str:
return_msg += "nnet_directory is not of type string, is type {}".format(
type(nnet_directory))
return {
RDK.success: RC.input_validation,
RDK.return_msg: return_msg,
RDK.debug_data: debug_data
}
else:
nnet_directory = KaldiNnetDecoder.CV_default_nnet_directory
if transcription_directory is not None:
if type(transcription_directory) is not str:
return_msg += "transcription_directory is not of type string, is type {}".format(
type(transcription_directory))
return {
RDK.success: RC.input_validation,
RDK.return_msg: return_msg,
RDK.debug_data: debug_data
}
else:
transcription_directory = KaldiNnetDecoder.CV_default_transcription_directory
##</end> input validation
## feats and ivector rspec creation
feats = (
"ark:compute-mfcc-feats --config={0}/conf/mfcc.conf scp:{1}/wav.scp ark:- |"
).format(nnet_directory, transcription_directory)
ivectors = (
"ark:compute-mfcc-feats --config={0}/conf/mfcc.conf scp:{1}/wav.scp ark:- |"
"ivector-extract-online2 --config={0}/conf/ivector_extractor.conf ark:{1}/spk2utt ark:- ark:- |"
).format(nnet_directory, transcription_directory)
##</end> feats and ivector rspec creation
## asr creation
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = 13
decoder_opts.max_active = 7000
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.acoustic_scale = 1.0
decodable_opts.frame_subsampling_factor = 3
decodable_opts.frames_per_chunk = 150
asr = NnetLatticeFasterRecognizer.from_files(
"{}/final.mdl".format(nnet_directory),
"{}/graph/HCLG.fst".format(nnet_directory),
"{}/graph/words.txt".format(nnet_directory),
decoder_opts=decoder_opts,
decodable_opts=decodable_opts)
##</end> asr creation
self.IV_feats = feats
self.IV_ivectors = ivectors
self.IV_asr = asr
self.IV_is_ready = True
return {
RDK.success: RC.success,
RDK.return_msg: return_msg,
RDK.debug_data: debug_data
}
def asr(filenameS_hash, filenameS, asr_beamsize=13, asr_max_active=8000):
models_dir = "models/"
# Read yaml File
config_file = "models/kaldi_tuda_de_nnet3_chain2.yaml"
with open(config_file, 'r') as stream:
model_yaml = yaml.safe_load(stream)
decoder_yaml_opts = model_yaml['decoder']
scp_filename = "tmp/%s.scp" % filenameS_hash
wav_filename = "tmp/%s.wav" % filenameS_hash
spk2utt_filename = "tmp/%s_spk2utt" % filenameS_hash
# write scp file
with open(scp_filename, 'w') as scp_file:
scp_file.write("%s tmp/%s.wav\n" % (filenameS_hash, filenameS_hash))
# write scp file
with open(spk2utt_filename, 'w') as scp_file:
scp_file.write("%s %s\n" % (filenameS_hash, filenameS_hash))
# use ffmpeg to convert the input media file (any format!) to 16 kHz wav mono
(
ffmpeg
.input(filename)
.output("tmp/%s.wav" % filenameS_hash, acodec='pcm_s16le', ac=1, ar='16k')
.overwrite_output()
.run()
)
# Construct recognizer
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = asr_beamsize
decoder_opts.max_active = asr_max_active
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.acoustic_scale = 1.0
decodable_opts.frame_subsampling_factor = 3
decodable_opts.frames_per_chunk = 150
asr = NnetLatticeFasterRecognizer.from_files(
models_dir + decoder_yaml_opts["model"],
models_dir + decoder_yaml_opts["fst"],
models_dir + decoder_yaml_opts["word-syms"],
decoder_opts=decoder_opts, decodable_opts=decodable_opts)
# Construct symbol table
symbols = SymbolTable.read_text(models_dir + decoder_yaml_opts["word-syms"])
phi_label = symbols.find_index("#0")
# Define feature pipelines as Kaldi rspecifiers
feats_rspec = ("ark:compute-mfcc-feats --config=%s scp:" + scp_filename + " ark:- |") % \
(models_dir + decoder_yaml_opts["mfcc-config"])
ivectors_rspec = (
("ark:compute-mfcc-feats --config=%s scp:" + scp_filename + " ark:-"
+ " | ivector-extract-online2 --config=%s ark:" + spk2utt_filename + " ark:- ark:- |") %
((models_dir + decoder_yaml_opts["mfcc-config"]),
(models_dir + decoder_yaml_opts["ivector-extraction-config"]))
)
did_decode = False
# Decode wav files
with SequentialMatrixReader(feats_rspec) as f, \
SequentialMatrixReader(ivectors_rspec) as i:
for (fkey, feats), (ikey, ivectors) in zip(f, i):
did_decode = True
assert (fkey == ikey)
out = asr.decode((feats, ivectors))
best_path = functions.compact_lattice_shortest_path(out["lattice"])
words, _, _ = get_linear_symbol_sequence(shortestpath(best_path))
timing = functions.compact_lattice_to_word_alignment(best_path)
assert(did_decode)
# Maps words to the numbers
words = indices_to_symbols(symbols, timing[0])
# Creates the datastructure (Word, begin(Frames), end(Frames))
vtt = list(map(list, zip(words, timing[1], timing[2])))
# Cleanup tmp files
print('removing tmp file:', scp_filename)
os.remove(scp_filename)
print('removing tmp file:', wav_filename)
os.remove(wav_filename)
print('removing tmp file:', spk2utt_filename)
os.remove(spk2utt_filename)
return vtt, words
# Reading from config file
parser = ConfigParser()
parser.read('sad_model.conf')
samplerate = int(parser.get('AUDIO', 'SAMPLE_RATE'))
n_channels = int(parser.get('AUDIO', 'N_CHANNELS'))
encoding = parser.get('AUDIO', 'ENCODING')
sad_final_raw = parser.get('SAD', 'FINAL_RAW')
post_output_vec = parser.get('SAD', 'POST_OUTPUT_VEC')
# Construct SAD
model = NnetSAD.read_model(sad_final_raw)
post = NnetSAD.read_average_posteriors(post_output_vec)
transform = NnetSAD.make_sad_transform(post)
graph = NnetSAD.make_sad_graph(min_silence_duration=0.1)
decodable_opts = NnetSimpleComputationOptions()
decodable_opts.extra_left_context = 79
decodable_opts.extra_right_context = 21
decodable_opts.extra_left_context_initial = 0
decodable_opts.extra_right_context_final = 0
decodable_opts.frames_per_chunk = 150
decodable_opts.acoustic_scale = 0.3
sad = NnetSAD(model, transform, graph, decodable_opts=decodable_opts)
seg = SegmentationProcessor(target_labels=[2],
min_segment_dur=3,
max_merged_segment_dur=3.84)
def convert_to_wav(path):
'''
1. Converts audio file in other formats to wav format