def main():
# ------------- Parse arguments from command line ----------------------
# 1. Add a discription of this program
args.discribe("This program is used to train triphone GMM-HMM model")
# 2. Add options
args.add("--expDir",
abbr="-e",
dtype=str,
default="exp",
discription="The data and output path of current experiment.")
args.add("--splice",
abbr="-c",
dtype=int,
default=3,
discription="How many left-right frames to splice.")
args.add("--numIters",
abbr="-n",
dtype=int,
default=35,
discription="How many iterations to train.")
args.add("--maxIterInc",
abbr="-m",
dtype=int,
default=25,
discription="The final iteration of increasing gaussians.")
args.add("--realignIter",
abbr="-r",
dtype=int,
default=[10, 20, 30],
discription="The iteration to realign feature.")
args.add("--fmllrIter",
abbr="-f",
dtype=int,
default=[2, 4, 6, 12],
discription="The iteration to estimate fmllr matrix.")
args.add("--order",
abbr="-o",
dtype=int,
default=6,
discription="Which N-grams model to use.")
args.add("--beam",
abbr="-b",
dtype=int,
default=13,
discription="Decode beam size.")
args.add("--latBeam",
abbr="-l",
dtype=int,
default=6,
discription="Lattice beam size.")
args.add("--acwt",
abbr="-a",
dtype=float,
default=0.083333,
discription="Acoustic model weight.")
args.add(
"--parallel",
abbr="-p",
dtype=int,
default=4,
minV=1,
maxV=10,
discription=
"The number of parallel process to compute feature of train dataset.")
args.add("--skipTrain",
abbr="-s",
dtype=bool,
default=False,
discription="If True, skip training. Do decoding only.")
# 3. Then start to parse arguments.
args.parse()
# 4. Take a backup of arguments
argsLogFile = os.path.join(args.expDir, "conf", "train_sat.args")
args.save(argsLogFile)
if not args.skipTrain:
# ------------- Prepare feature and previous alignment for training ----------------------
# 1. Load the feature for training
print(f"Load MFCC+CMVN feature.")
feat = exkaldi.load_index_table(
os.path.join(args.expDir, "mfcc", "train", "mfcc_cmvn.ark"))
print(f"Splice {args.splice} frames.")
originalFeat = exkaldi.splice_feature(feat,
left=args.splice,
right=args.splice,
outFile=os.path.join(
args.expDir, "train_delta",
"mfcc_cmvn_splice.ark"))
print(f"Transform LDA feature")
ldaFeat = exkaldi.transform_feat(
feat=originalFeat,
matFile=os.path.join(args.expDir, "train_lda_mllt", "trans.mat"),
outFile=os.path.join(args.expDir, "train_sat", "lda_feat.ark"),
)
del originalFeat
# 2. Load previous alignment and lexicons
ali = exkaldi.load_index_table(os.path.join(args.expDir,
"train_lda_mllt",
"*final.ali"),
useSuffix="ark")
lexicons = exkaldi.load_lex(
os.path.join(args.expDir, "dict", "lexicons.lex"))
# 3. Estimate the primary fMLLR transform matrix
print("Estiminate the primary fMLLR transform matrixs")
fmllrTransMat = exkaldi.hmm.estimate_fMLLR_matrix(
aliOrLat=ali,
lexicons=lexicons,
aliHmm=os.path.join(args.expDir, "train_lda_mllt", "final.mdl"),
feat=ldaFeat,
spk2utt=os.path.join(args.expDir, "data", "train", "spk2utt"),
outFile=os.path.join(args.expDir, "train_sat", "trans.ark"),
)
print("Transform feature")
fmllrFeat = exkaldi.use_fmllr(
ldaFeat,
fmllrTransMat,
utt2spk=os.path.join("exp", "data", "train", "utt2spk"),
outFile=os.path.join(args.expDir, "train_sat", "fmllr_feat.ark"),
)
# -------------- Build the decision tree ------------------------
print("Start build a tree")
tree = exkaldi.hmm.DecisionTree(lexicons=lexicons,
contextWidth=3,
centralPosition=1)
tree.train(
feat=fmllrFeat,
hmm=os.path.join(args.expDir, "train_lda_mllt", "final.mdl"),
ali=ali,
topoFile=os.path.join(args.expDir, "dict", "topo"),
numLeaves=2500,
tempDir=os.path.join(args.expDir, "train_sat"),
)
tree.save(os.path.join(args.expDir, "train_sat", "tree"))
print(f"Build tree done.")
del fmllrFeat
# ------------- Start training ----------------------
# 1. Initialize a monophone HMM object
print("Initialize a triphone HMM object")
model = exkaldi.hmm.TriphoneHMM(lexicons=lexicons)
model.initialize(
tree=tree,
topoFile=os.path.join(args.expDir, "dict", "topo"),
treeStatsFile=os.path.join(args.expDir, "train_sat",
"treeStats.acc"),
)
print(f"Initialized a monophone HMM-GMM model: {model.info}.")
# 2. convert the previous alignment
print(f"Transform the alignment")
newAli = exkaldi.hmm.convert_alignment(
ali=ali,
originHmm=os.path.join(args.expDir, "train_lda_mllt", "final.mdl"),
targetHmm=model,
tree=tree,
outFile=os.path.join(args.expDir, "train_sat", "initial.ali"),
)
# 2. Split data for parallel training
transcription = exkaldi.load_transcription(
os.path.join(args.expDir, "data", "train", "text"))
transcription = transcription.sort()
if args.parallel > 1:
# split feature
ldaFeat = ldaFeat.sort(by="utt").subset(chunks=args.parallel)
# split transcription depending on utterance IDs of each feat
tempTrans = []
tempAli = []
tempFmllrMat = []
for f in ldaFeat:
tempTrans.append(transcription.subset(keys=f.utts))
tempAli.append(newAli.subset(keys=f.utts))
spks = exkaldi.utt_to_spk(f.utts,
utt2spk=os.path.join(
args.expDir, "data", "train",
"utt2spk"))
tempFmllrMat.append(fmllrTransMat.subset(keys=spks))
transcription = tempTrans
newAli = tempAli
fmllrTransMat = tempFmllrMat
# 3. Train
print("Train the triphone model")
model.train(
ldaFeat,
transcription,
os.path.join(args.expDir, "dict", "L.fst"),
tree,
tempDir=os.path.join(args.expDir, "train_sat"),
initialAli=newAli,
fmllrTransMat=fmllrTransMat,
spk2utt=os.path.join(args.expDir, "data", "train", "spk2utt"),
utt2spk=os.path.join(args.expDir, "data", "train", "utt2spk"),
numIters=args.numIters,
maxIterInc=args.maxIterInc,
totgauss=15000,
realignIter=args.realignIter,
fmllrIter=args.fmllrIter,
boostSilence=1.0,
power=0.2,
fmllrSilWt=0.0,
)
print(model.info)
del ldaFeat
del fmllrTransMat
del newAli
else:
declare.is_file(os.path.join(args.expDir, "train_sat", "final.mdl"))
declare.is_file(os.path.join(args.expDir, "train_sat", "tree"))
model = exkaldi.load_hmm(
os.path.join(args.expDir, "train_sat", "final.mdl"))
tree = exkaldi.load_tree(os.path.join(args.expDir, "train_sat",
"tree"))
# ------------- Compile WFST training ----------------------
# Make a WFST decoding graph
make_WFST_graph(
outDir=os.path.join(args.expDir, "train_sat", "graph"),
hmm=model,
tree=tree,
)
# Decode test data
GMM_decode_fmllr_and_score(
outDir=os.path.join(args.expDir, "train_sat",
f"decode_{args.order}grams"),
hmm=model,
HCLGfile=os.path.join(args.expDir, "train_sat", "graph",
f"HCLG.{args.order}.fst"),
tansformMatFile=os.path.join(args.expDir, "train_lda_mllt",
"trans.mat"),
)
def train():
# ------------- Prepare data for dnn training ----------------------
stamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
outDir = f"dnn_exp/out_{stamp}"
exkaldi.utils.make_dependent_dirs(outDir, pathIsFile=False)
args.save( os.path.join(outDir,"conf") )
#------------------------ Training and Validation dataset-----------------------------
hmm = exkaldi.load_hmm(f"{args.root}/exp/tri3b_ali_train_clean_5/final.mdl")
pdfDim = hmm.info.pdfs
del hmm
print('Prepare Data Iterator...')
# Prepare fMLLR feature files
featDim, trainDataset = prepare_data()
traindataLen = len(trainDataset)
train_gen = tf.data.Dataset.from_generator(
lambda: make_generator(trainDataset),
(tf.float32, tf.int32),
(tf.TensorShape([featDim,]), tf.TensorShape([])),
).batch(args.batchSize).prefetch(3)
steps_per_epoch = traindataLen//args.batchSize
featDim, devDataset = prepare_data(training=False)
devdataLen = len(devDataset)
dev_gen = tf.data.Dataset.from_generator(
lambda: make_generator(devDataset),
(tf.float32, tf.int32),
(tf.TensorShape([featDim,]), tf.TensorShape([])),
).batch(args.batchSize).prefetch(3)
validation_steps = devdataLen//args.batchSize
#------------------------ Train Step -----------------------------
model = make_DNN_acoustic_model(featDim,pdfDim)
#model.summary()
model.compile(
loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics = keras.metrics.SparseCategoricalAccuracy(),
optimizer = keras.optimizers.SGD(0.08,momentum=0.0),
)
def lrScheduler(epoch):
if epoch > 25:
return 0.001
elif epoch > 22:
return 0.0025
elif epoch > 19:
return 0.005
elif epoch > 17:
return 0.01
elif epoch > 15:
return 0.02
elif epoch > 10:
return 0.04
else:
return 0.08
model.fit(
x = train_gen,
steps_per_epoch=steps_per_epoch,
epochs=args.epoch,
validation_data=dev_gen,
validation_steps=validation_steps,
verbose=1,
initial_epoch=0,
callbacks=[
keras.callbacks.EarlyStopping(patience=5, verbose=1),
keras.callbacks.TensorBoard(log_dir=outDir),
keras.callbacks.LearningRateScheduler(lrScheduler),
ModelSaver(model,outDir),
],
)
def main():
# ------------- Parse arguments from command line ----------------------
# 1. Add a discription of this program
args.describe("This program is used to train monophone GMM-HMM model")
# 2. Add options
args.add("--expDir",
abbr="-e",
dtype=str,
default="exp",
discription="The data and output path of current experiment.")
args.add("--delta",
abbr="-d",
dtype=int,
default=2,
discription="Add n-order to feature.")
args.add("--numIters",
abbr="-n",
dtype=int,
default=40,
discription="How many iterations to train.")
args.add("--maxIterInc",
abbr="-m",
dtype=int,
default=30,
discription="The final iteration of increasing gaussians.")
args.add("--realignIter",
abbr="-r",
dtype=int,
default=[
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16, 18, 20, 23, 26, 29,
32, 35, 38
],
discription="the iteration to realign feature.")
args.add("--order",
abbr="-o",
dtype=int,
default=6,
minV=1,
maxV=6,
discription="Which N-grams model to use.")
args.add("--beam",
abbr="-b",
dtype=int,
default=13,
discription="Decode beam size.")
args.add("--latBeam",
abbr="-l",
dtype=int,
default=6,
discription="Lattice beam size.")
args.add("--acwt",
abbr="-a",
dtype=float,
default=0.083333,
discription="Acoustic model weight.")
args.add(
"--parallel",
abbr="-p",
dtype=int,
default=4,
minV=1,
maxV=10,
discription=
"The number of parallel process to compute feature of train dataset.")
args.add("--skipTrain",
abbr="-s",
dtype=bool,
default=False,
discription="If True, skip training. Do decoding only.")
# 3. Then start to parse arguments.
args.parse()
# 4. Take a backup of arguments
args.print_args() # print arguments to display
argsLogFile = os.path.join(args.expDir, "conf", "train_mono.args")
args.save(argsLogFile)
if not args.skipTrain:
# ------------- Prepare feature for training ----------------------
# 1. Load the feature for training (We use the index table format)
feat = exkaldi.load_index_table(
os.path.join(args.expDir, "mfcc", "train", "mfcc_cmvn.ark"))
print(f"Load MFCC+CMVN feature.")
feat = exkaldi.add_delta(feat,
order=args.delta,
outFile=os.path.join(args.expDir,
"train_mono",
"mfcc_cmvn_delta.ark"))
print(f"Add {args.delta}-order deltas.")
# 2. Load lexicon bank
lexicons = exkaldi.load_lex(
os.path.join(args.expDir, "dict", "lexicons.lex"))
print(f"Restorage lexicon bank.")
# ------------- Start training ----------------------
# 1. Initialize a monophone HMM object
model = exkaldi.hmm.MonophoneHMM(lexicons=lexicons, name="mono")
model.initialize(feat=feat,
topoFile=os.path.join(args.expDir, "dict", "topo"))
print(f"Initialized a monophone HMM-GMM model: {model.info}.")
# 2. Split data for parallel training
transcription = exkaldi.load_transcription(
os.path.join(args.expDir, "data", "train", "text"))
transcription = transcription.sort()
if args.parallel > 1:
# split feature
feat = feat.sort(by="utt").subset(chunks=args.parallel)
# split transcription depending on utterance IDs of each feature
temp = []
for f in feat:
temp.append(transcription.subset(keys=f.utts))
transcription = temp
# 3. Train
model.train(
feat,
transcription,
LFile=os.path.join(args.expDir, "dict", "L.fst"),
tempDir=os.path.join(args.expDir, "train_mono"),
numIters=args.numIters,
maxIterInc=args.maxIterInc,
totgauss=1000,
realignIter=args.realignIter,
boostSilence=1.0,
)
print(model.info)
# Save the tree
model.tree.save(os.path.join(args.expDir, "train_mono", "tree"))
print(f"Tree has been saved.")
# 4. Realign with boostSilence 1.25
print("Realign the training feature (boost silence = 1.25)")
trainGraphFiles = exkaldi.utils.list_files(
os.path.join(args.expDir, "train_mono", "*train_graph"))
model.align(
feat,
trainGraphFile=
trainGraphFiles, # train graphs have been generated in the train step.
boostSilence=1.25, #1.5
outFile=os.path.join(args.expDir, "train_mono", "final.ali"))
del feat
print("Save the new alignment done.")
tree = model.tree
else:
declare.is_file(os.path.join(args.expDir, "train_mono", "final.mdl"))
declare.is_file(os.path.join(args.expDir, "train_mono", "tree"))
model = exkaldi.load_hmm(
os.path.join(args.expDir, "train_mono", "final.mdl"))
tree = exkaldi.load_tree(
os.path.join(args.expDir, "train_mono", "tree"))
# ------------- Compile WFST training ----------------------
# Make a WFST decoding graph
make_WFST_graph(
outDir=os.path.join(args.expDir, "train_mono", "graph"),
hmm=model,
tree=tree,
)
# Decode test data
GMM_decode_mfcc_and_score(
outDir=os.path.join(args.expDir, "train_mono",
f"decode_{args.order}grams"),
hmm=model,
HCLGfile=os.path.join(args.expDir, "train_mono", "graph",
f"HCLG.{args.order}.fst"),
)
def compute_dev_wer():
flag = "dev_clean_2"
featsFile = f"{args.root}/{args.feat}/raw_{args.feat}_{flag}.*.ark"
feats = exkaldi.load_feat(featsFile)
if args.cmn:
print("Use cmvn...")
cmvnFile = f"{args.root}/{args.feat}/cmvn_{flag}.ark"
cmvn = exkaldi.load_cmvn(cmvnFile)
feats = exkaldi.use_cmvn(feats,cmvn,utt2spk=f"{args.root}/data/{flag}/utt2spk")
del cmvn
if args.delta > 0:
print("Add delta...")
feats = feats.add_delta(args.delta)
if args.splice > 0:
print("Splice feature...")
feats = feats.splice(args.splice)
feats = feats.to_numpy()
featDim = feats.dim
hmm = exkaldi.load_hmm(f"{args.root}/exp/tri3b_ali_train_clean_5/final.mdl")
pdfDim = hmm.info.pdfs
phoneDim = hmm.info.phones
del hmm
print("featDim:",featDim,"pdfDim:",pdfDim,"phoneDim:",phoneDim)
minWER = None
try:
for ModelPathID in range(args.epoch,0,-1):
#ModelPathID = args.epoch
ModelPath = f"{args.testModelDir}/model_ep{ModelPathID}.h5"
if not os.path.isfile(ModelPath):
continue
print("Use Model:",ModelPath)
decodeOut = ModelPath[:-3]
exkaldi.utils.make_dependent_dirs(decodeOut,pathIsFile=False)
model = make_DNN_acoustic_model(featDim,pdfDim)
model.load_weights(ModelPath)
print("Forward...")
result = {}
for uttID in feats.keys():
pdfP = model(feats[uttID],training=False)
result[uttID] = exkaldi.nn.log_softmax(pdfP.numpy(),axis=1)
amp = exkaldi.load_prob(result)
hmmFile = f"{args.root}/exp/tri3b_ali_dev_clean_2/final.mdl"
HCLGFile = f"{args.root}/exp/tri3b/graph_tgsmall/HCLG.fst"
table = f"{args.root}/exp/tri3b/graph_tgsmall/words.txt"
trans = f"{args.root}/data/dev_clean_2/text"
print("Decoding...")
lat = exkaldi.decode.wfst.nn_decode(
prob=amp.subset(chunks=4),
hmm=hmmFile,
HCLGFile=HCLGFile,
symbolTable=table,
beam=10,
latBeam=8,
acwt=0.1,
minActive=200,
maxActive=7000,
outFile=os.path.join(decodeOut,"lat")
)
lat = exkaldi.merge_archives(lat)
print("Scoring...")
for LMWT in range(1,10,1):
#newLat = lat.add_penalty(penalty)
result = lat.get_1best(table,hmmFile,lmwt=LMWT,acwt=0.1,phoneLevel=False)
result = exkaldi.hmm.transcription_from_int(result,table)
result.save( os.path.join(decodeOut,f"trans.{LMWT}") )
score = exkaldi.decode.score.wer(ref=trans,hyp=result,mode="present")
print("LMWT: ",LMWT ,"WER: ",score.WER)
if minWER == None or score.WER < minWER[0]:
minWER = (score.WER, LMWT, ModelPath)
finally:
if minWER is not None:
werOut = os.path.basename(decodeOut)
print("Best WER:",minWER)
with open(f"{args.testModelDir}/best_wer","w") as fw:
fw.write(str(minWER))
def main():
# ------------- Parse arguments from command line ----------------------
# 1. Add a discription of this program
args.discribe("This program is used to train triphone GMM-HMM model")
# 2. Add options
args.add("--expDir",
abbr="-e",
dtype=str,
default="exp",
discription="The data and output path of current experiment.")
args.add("--delta",
abbr="-d",
dtype=int,
default=2,
discription="Add n-order to feature.")
args.add("--numIters",
abbr="-n",
dtype=int,
default=35,
discription="How many iterations to train.")
args.add("--maxIterInc",
abbr="-m",
dtype=int,
default=25,
discription="The final iteration of increasing gaussians.")
args.add("--realignIter",
abbr="-r",
dtype=int,
default=[10, 20, 30],
discription="the iteration to realign feature.")
args.add("--order",
abbr="-o",
dtype=int,
default=6,
discription="Which N-grams model to use.")
args.add("--beam",
abbr="-b",
dtype=int,
default=13,
discription="Decode beam size.")
args.add("--latBeam",
abbr="-l",
dtype=int,
default=6,
discription="Lattice beam size.")
args.add("--acwt",
abbr="-a",
dtype=float,
default=0.083333,
discription="Acoustic model weight.")
args.add(
"--parallel",
abbr="-p",
dtype=int,
default=4,
minV=1,
maxV=10,
discription=
"The number of parallel process to compute feature of train dataset.")
args.add("--skipTrain",
abbr="-s",
dtype=bool,
default=False,
discription="If True, skip training. Do decoding only.")
# 3. Then start to parse arguments.
args.parse()
# 4. Take a backup of arguments
argsLogFile = os.path.join(args.expDir, "conf", "train_delta.args")
args.save(argsLogFile)
if not args.skipTrain:
# ------------- Prepare feature and previous alignment for training ----------------------
# 1. Load the feature for training
feat = exkaldi.load_index_table(
os.path.join(args.expDir, "mfcc", "train", "mfcc_cmvn.ark"))
print(f"Load MFCC+CMVN feature.")
feat = exkaldi.add_delta(feat,
order=args.delta,
outFile=os.path.join(args.expDir,
"train_delta",
"mfcc_cmvn_delta.ark"))
print(f"Add {args.delta}-order deltas.")
# 2. Load lexicon bank
lexicons = exkaldi.load_lex(
os.path.join(args.expDir, "dict", "lexicons.lex"))
print(f"Restorage lexicon bank.")
# 3. Load previous alignment
ali = exkaldi.load_index_table(os.path.join(args.expDir, "train_mono",
"*final.ali"),
useSuffix="ark")
# -------------- Build the decision tree ------------------------
print("Start build a tree")
tree = exkaldi.hmm.DecisionTree(lexicons=lexicons,
contextWidth=3,
centralPosition=1)
tree.train(
feat=feat,
hmm=os.path.join(args.expDir, "train_mono", "final.mdl"),
ali=ali,
topoFile=os.path.join(args.expDir, "dict", "topo"),
numLeaves=2500,
tempDir=os.path.join(args.expDir, "train_delta"),
)
print(f"Build tree done.")
# ------------- Start training ----------------------
# 1. Initialize a monophone HMM object
model = exkaldi.hmm.TriphoneHMM(lexicons=lexicons, name="mono")
model.initialize(
tree=tree,
topoFile=os.path.join(args.expDir, "dict", "topo"),
treeStatsFile=os.path.join(args.expDir, "train_delta",
"treeStats.acc"),
)
print(f"Initialized a monophone HMM-GMM model: {model.info}.")
# 2. convert the previous alignment
print(f"Transform the alignment")
newAli = exkaldi.hmm.convert_alignment(
ali=ali,
originHmm=os.path.join("exp", "train_mono", "final.mdl"),
targetHmm=model,
tree=tree,
outFile=os.path.join(args.expDir, "train_delta", "initial.ali"),
)
# 2. Split data for parallel training
transcription = exkaldi.load_transcription(
os.path.join(args.expDir, "data", "train", "text"))
transcription = transcription.sort()
if args.parallel > 1:
# split feature
feat = feat.sort(by="utt").subset(chunks=args.parallel)
# split transcription depending on utterance IDs of each feat
tempTrans = []
tempAli = []
for f in feat:
tempTrans.append(transcription.subset(keys=f.utts))
tempAli.append(newAli.subset(keys=f.utts))
transcription = tempTrans
newAli = tempAli
# 3. Train
print("Train the triphone model")
model.train(
feat,
transcription,
os.path.join("exp", "dict", "L.fst"),
tree,
tempDir=os.path.join(args.expDir, "train_delta"),
initialAli=newAli,
numIters=args.numIters,
maxIterInc=args.maxIterInc,
totgauss=15000,
realignIter=args.realignIter,
boostSilence=1.0,
)
print(model.info)
# Save the tree
model.tree.save(os.path.join(args.expDir, "train_delta", "tree"))
print(f"Tree has been saved.")
del feat
else:
declare.is_file(os.path.join(args.expDir, "train_delta", "final.mdl"))
declare.is_file(os.path.join(args.expDir, "train_delta", "tree"))
model = exkaldi.load_hmm(
os.path.join(args.expDir, "train_delta", "final.mdl"))
tree = exkaldi.load_tree(
os.path.join(args.expDir, "train_delta", "tree"))
# ------------- Compile WFST training ----------------------
# Make a WFST decoding graph
make_WFST_graph(
outDir=os.path.join(args.expDir, "train_delta", "graph"),
hmm=model,
tree=tree,
)
# Decode test data
GMM_decode_mfcc_and_score(
outDir=os.path.join(args.expDir, "train_delta",
f"decode_{args.order}grams"),
hmm=model,
HCLGfile=os.path.join(args.expDir, "train_delta", "graph",
f"HCLG.{args.order}.fst"),
)