def mixture_feat_extractor_test():
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=None,
)
cutter = stream.ElementFrameCutter(
width=400,
shift=160,
)
extractor = feature.MixtureExtractor(
frameDim=400,
batchSize=100,
mixType=["mfcc", "fbank"],
useEnergyForFbank=False,
useEnergyForMfcc=False,
)
reader.start()
cutter.start(inPIPE=reader.outPIPE)
extractor.start(inPIPE=cutter.outPIPE)
extractor.wait()
print(extractor.outPIPE.size())
pac = extractor.outPIPE.get()
print(pac.data.shape)
def feat_extractor_test():
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=None,
)
cutter = stream.ElementFrameCutter(
width=400,
shift=160,
)
extractor = feature.MfccExtractor(
batchSize=100,
useEnergy=False,
)
reader.start()
cutter.start(inPIPE=reader.outPIPE)
extractor.start(inPIPE=cutter.outPIPE)
extractor.wait()
print(extractor.outPIPE.size())
def stream_reader_test():
vad = None #stream.WebrtcVADetector()
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=vad,
)
reader.start()
reader.wait()
print(reader.outPIPE.size())
def feat_estimator_test():
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=None,
)
cutter = stream.ElementFrameCutter(
width=400,
shift=160,
)
extractor = feature.MfccExtractor(
batchSize=100,
useEnergy=False,
)
processor = feature.FeatureProcessor(
featDim=13,
delta=2,
spliceLeft=10,
spliceRight=10,
cmvNormalizer=feature.FrameSlideCMVNormalizer(),
)
left = 5
right = 5
estimator = decode.AcousticEstimator(
featDim=819,
batchSize=100,
applySoftmax=False,
applyLog=False,
leftContext=left,
rightContext=right,
)
estimator.acoustic_function = lambda x: x[left:-right].copy()
reader.start()
cutter.start(inPIPE=reader.outPIPE)
extractor.start(inPIPE=cutter.outPIPE)
processor.start(inPIPE=extractor.outPIPE)
estimator.start(inPIPE=processor.outPIPE)
estimator.wait()
print(estimator.outPIPE.size())
def cutter_test():
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=None,
)
cutter = stream.ElementFrameCutter(
width=400,
shift=160,
)
reader.start()
cutter.start(inPIPE=reader.outPIPE)
cutter.wait()
print(cutter.outPIPE.size())
def send_value_packets():
wavPath = "../examples/84-121550-0000.wav"
assert os.path.isfile(wavPath), f"No such file: {wavPath}"
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=None,
)
sender = transmit.PacketSender(
thost="192.168.1.11",
tport=9509,
batchSize=1024,
)
sender.encode_function = transmit.encode_value_packet
reader.start()
sender.start(inPIPE=reader.outPIPE)
sender.wait()
def feat_processor_test():
reader = stream.StreamReader(
waveFile=wavPath,
chunkSize=480,
simulate=False,
vaDetector=None,
)
cutter = stream.ElementFrameCutter(
width=400,
shift=160,
)
extractor = feature.MfccExtractor(
batchSize=100,
useEnergy=False,
)
processor = feature.FeatureProcessor(
featDim=13,
delta=2,
spliceLeft=10,
spliceRight=10,
cmvNormalizer=feature.FrameSlideCMVNormalizer(),
)
reader.start()
cutter.start(inPIPE=reader.outPIPE)
extractor.start(inPIPE=cutter.outPIPE)
processor.start(inPIPE=extractor.outPIPE)
processor.wait()
print(processor.outPIPE.size())
pac = processor.outPIPE.get()
print(pac.data.shape)
rootDir = f"{KALDI_ROOT}/egs/mini_librispeech/s5/exp"
words = f"{rootDir}/tri3b/graph_tgsmall/words.txt"
hmm = f"{rootDir}/tri3b_ali_train_clean_5/final.mdl"
HCLG = f"{rootDir}/tri3b/graph_tgsmall/HCLG.fst"
pdfDim = decode.get_pdf_dim(hmm)
kerasmodel = make_DNN_acoustic_model(featDim, pdfDim)
kerasmodel.load_weights(kerasModel)
##########################
# Define components
##########################
# 1. Create a stream reader to read realtime stream from audio file
reader = stream.StreamReader(waveFile, simulate=True)
# 2. Cutter to cut frame
cutter = stream.ElementFrameCutter(width=400, shift=160)
# 3. MFCC feature extracting
extractor = feature.MfccExtractor(
frameDim=400,
batchSize=100,
useEnergy=False,
)
# 4. processing feature
processor = feature.FeatureProcessor(
featDim=13,
batchSize=100,
delta=delta,
spliceLeft=spliceLeft,
spliceRight=spliceRight,
rHostIP = "192.168.1.11"
rHostPort = 9509
bHostPort = 9510
assert os.path.isfile(waveFile), f"No such file: {waveFile}"
##########################
# Define components
##########################
# 1. Create a stream reader to read realtime stream from audio file
vad = stream.WebrtcVADetector()
reader = stream.StreamReader(
waveFile=waveFile,
chunkSize=480,
simulate=True,
vaDetector=vad,
)
# 2. Send packets to remote host
sender = transmit.PacketSender(
thost=rHostIP,
tport=rHostPort,
batchSize=100,
)
sender.encode_function = transmit.encode_value_packet
# 3. Receive packets
receiver = transmit.PacketReceiver(bport=bHostPort)
featDim = (13 * (delta + 1)) * (spliceLeft + 1 + spliceRight)
##########################
# Load DNN acoustic model
##########################
pdfDim = decode.get_pdf_dim(hmm)
kerasmodel = make_DNN_acoustic_model(featDim, pdfDim)
kerasmodel.load_weights(kerasModel)
##########################
# Define components
##########################
# 1. Create a stream reader to read realtime stream from audio file
reader = stream.StreamReader(waveFile, simulate=False)
# 2. Cutter to cut frame
cutter = stream.ElementFrameCutter(batchSize=50, width=400, shift=160)
# 3. MFCC feature extracting
extractor = feature.MfccExtractor(useEnergy=False, )
# 4. processing feature
processor = feature.MatrixFeatureProcessor(
delta=delta,
spliceLeft=spliceLeft,
spliceRight=spliceRight,
cmvNormalizer=feature.FrameSlideCMVNormalizer(),
)
# 5. acoustic probability computer
def keras_compute(feats):