def prepare_DNN_data():
print("Start to prepare data for DNN training")
assert os.path.isdir(f"{args.expDir}/train_sat"
), "Please run previous programs up to SAT training."
# Lexicons and Gmm-Hmm model
lexicons = exkaldi.load_lex(f"{args.expDir}/dict/lexicons.lex")
hmm = f"{args.expDir}/train_sat/final.mdl"
tree = f"{args.expDir}/train_sat/tree"
for Name in ["train", "dev", "test"]:
exkaldi.utils.make_dependent_dirs(
f"{args.expDir}/train_dnn/data/{Name}", pathIsFile=False)
# Make LDA feature
print(f"Make LDA feature for '{Name}'")
feat = exkaldi.load_feat(f"{args.expDir}/mfcc/{Name}/mfcc_cmvn.ark")
feat = feat.splice(left=args.LDAsplice, right=args.LDAsplice)
feat = exkaldi.transform_feat(
feat, matFile=f"{args.expDir}/train_lda_mllt/trans.mat")
# Compile the aligning graph
print(f"Compile aligning graph")
transInt = exkaldi.hmm.transcription_to_int(
transcription=f"{args.expDir}/data/{Name}/text",
symbolTable=lexicons("words"),
unkSymbol=lexicons("oov"),
)
graphFile = exkaldi.decode.wfst.compile_align_graph(
hmm,
tree,
transcription=transInt,
LFile=f"{args.expDir}/dict/L.fst",
outFile=f"{args.expDir}/train_dnn/data/{Name}/align_graph",
lexicons=lexicons,
)
# Align first time
print(f"Align the first time")
ali = exkaldi.decode.wfst.gmm_align(
hmm,
feat,
alignGraphFile=graphFile,
lexicons=lexicons,
)
# Estimate transform matrix
print(f"Estimate fMLLR transform matrix")
fmllrTransMat = exkaldi.hmm.estimate_fMLLR_matrix(
aliOrLat=ali,
lexicons=lexicons,
aliHmm=hmm,
feat=feat,
spk2utt=f"{args.expDir}/data/{Name}/spk2utt",
)
fmllrTransMat.save(f"{args.expDir}/train_dnn/data/{Name}/trans.ark")
# Transform feature
print(f"Transform feature")
feat = exkaldi.use_fmllr(
feat,
fmllrTransMat,
utt2spk=f"{args.expDir}/data/{Name}/utt2spk",
)
# Align second time with new feature
print(f"Align the second time")
ali = exkaldi.decode.wfst.gmm_align(
hmm,
feat,
alignGraphFile=graphFile,
lexicons=lexicons,
)
# Save alignment and feature
print(f"Save final fmllr feature and alignment")
feat.save(f"{args.expDir}/train_dnn/data/{Name}/fmllr.ark")
ali.save(f"{args.expDir}/train_dnn/data/{Name}/ali")
# Transform alignment
print(f"Generate pdf ID and phone ID alignment")
ali.to_numpy(
aliType="pdfID",
hmm=hmm).save(f"{args.expDir}/train_dnn/data/{Name}/pdfID.npy")
ali.to_numpy(
aliType="phoneID",
hmm=hmm).save(f"{args.expDir}/train_dnn/data/{Name}/phoneID.npy")
del ali
# Compute cmvn for fmllr feature
print(f"Compute the CMVN for fmllr feature")
cmvn = exkaldi.compute_cmvn_stats(
feat, spk2utt=f"{args.expDir}/data/{Name}/spk2utt")
cmvn.save(f"{args.expDir}/train_dnn/data/{Name}/cmvn_of_fmllr.ark")
del cmvn
del feat
# copy spk2utt utt2spk and text file
shutil.copyfile(f"{args.expDir}/data/{Name}/spk2utt",
f"{args.expDir}/train_dnn/data/{Name}/spk2utt")
shutil.copyfile(f"{args.expDir}/data/{Name}/utt2spk",
f"{args.expDir}/train_dnn/data/{Name}/utt2spk")
shutil.copyfile(f"{args.expDir}/data/{Name}/text",
f"{args.expDir}/train_dnn/data/{Name}/text")
transInt.save(f"{args.expDir}/data/{Name}/text.int")
print("Write feature and alignment dim information")
dims = exkaldi.ListTable()
feat = exkaldi.load_feat(f"{args.expDir}/train_dnn/data/test/fmllr.ark")
dims["fmllr"] = feat.dim
del feat
hmm = exkaldi.hmm.load_hmm(f"{args.expDir}/train_sat/final.mdl")
dims["phones"] = hmm.info.phones + 1
dims["pdfs"] = hmm.info.pdfs
del hmm
dims.save(f"{args.expDir}/train_dnn/data/dims")
def prepare_LSTM_data():
print("Start to prepare data for LSTM training")
declare.is_dir(f"{args.expDir}/train_dnn/prob", debug="Please run previous programs up to DNN training.")
# Lexicons and Gmm-Hmm model
lexicons = exkaldi.load_lex( f"{args.expDir}/dict/lexicons.lex" )
hmm = f"{args.expDir}/train_sat/final.mdl"
tree = f"{args.expDir}/train_sat/tree"
for Name in ["train", "dev", "test"]:
exkaldi.utils.make_dependent_dirs(f"{args.expDir}/train_lstm/data/{Name}", pathIsFile=False)
# Load feature
print(f"Make LDA feature for '{Name}'")
feat = exkaldi.load_feat( f"{args.expDir}/mfcc/{Name}/mfcc_cmvn.ark" )
feat = feat.splice(left=args.LDAsplice, right=args.LDAsplice)
feat = exkaldi.transform_feat(feat, matFile=f"{args.expDir}/train_lda_mllt/trans.mat" )
# Load probability for aligning( File has a large size, so we use index table. )
prob = exkaldi.load_index_table( f"{args.expDir}/train_dnn/prob/{Name}.ark" )
# Compile a aligning graph
print(f"Copy aligning graph from DNN resources")
shutil.copyfile( f"{args.expDir}/train_dnn/data/{Name}/align_graph",
f"{args.expDir}/train_lstm/data/{Name}/align_graph"
)
# Align
print("Align")
ali = exkaldi.decode.wfst.nn_align(
hmm,
prob,
alignGraphFile=f"{args.expDir}/train_lstm/data/{Name}/align_graph",
lexicons=lexicons,
outFile=f"{args.expDir}/train_lstm/data/{Name}/ali",
)
# Estimate transform matrix
print("Estimate transform matrix")
fmllrTransMat = exkaldi.hmm.estimate_fMLLR_matrix(
aliOrLat=ali,
lexicons=lexicons,
aliHmm=hmm,
feat=feat,
spk2utt=f"{args.expDir}/data/{Name}/spk2utt",
outFile=f"{args.expDir}/train_lstm/data/{Name}/trans.ark",
)
# Transform feature
print("Transform matrix")
feat = exkaldi.use_fmllr(
feat,
fmllrTransMat,
utt2spk=f"{args.expDir}/data/{Name}/utt2spk",
outFile=f"{args.expDir}/train_lstm/data/{Name}/fmllr.ark",
)
# Transform alignment (Because 'ali' is a index table object, we need fetch the alignment data in order to use the 'to_numpy' method.)
ali = ali.fetch(arkType="ali")
ali.to_numpy(aliType="pdfID",hmm=hmm).save( f"{args.expDir}/train_lstm/data/{Name}/pdfID.npy" )
ali.to_numpy(aliType="phoneID",hmm=hmm).save( f"{args.expDir}/train_lstm/data/{Name}/phoneID.npy" )
del ali
# Compute cmvn for fmllr feature
cmvn = exkaldi.compute_cmvn_stats(
feat,
spk2utt=f"{args.expDir}/data/{Name}/spk2utt",
outFile=f"{args.expDir}/train_lstm/data/{Name}/cmvn_of_fmllr.ark",
)
del cmvn
del feat
# copy spk2utt utt2spk and text file
shutil.copyfile( f"{args.expDir}/data/{Name}/spk2utt", f"{args.expDir}/train_lstm/data/{Name}/spk2utt")
shutil.copyfile( f"{args.expDir}/data/{Name}/utt2spk", f"{args.expDir}/train_lstm/data/{Name}/utt2spk")
shutil.copyfile( f"{args.expDir}/data/{Name}/text", f"{args.expDir}/train_lstm/data/{Name}/text" )
print("Write feature and alignment dim information")
dims = exkaldi.ListTable()
feat = exkaldi.load_feat( f"{args.expDir}/train_lstm/data/test/fmllr.ark" )
dims["fmllr"] = feat.dim
del feat
hmm = exkaldi.hmm.load_hmm( f"{args.expDir}/train_sat/final.mdl" )
dims["phones"] = hmm.info.phones + 1
dims["pdfs"] = hmm.info.pdfs
del hmm
dims.save( f"{args.expDir}/train_lstm/data/dims" )
def GMM_decode_fmllr_and_score(outDir, hmm, HCLGfile, tansformMatFile=None):
exkaldi.utils.make_dependent_dirs(outDir, pathIsFile=False)
lexicons = exkaldi.decode.graph.load_lex(
os.path.join("exp", "dict", "lexicons.lex"))
print(f"Load test feature.")
featFile = os.path.join("exp", "mfcc", "test", "mfcc_cmvn.ark")
feat = exkaldi.load_feat(featFile)
if tansformMatFile is None:
print("Feature type is delta")
feat = feat.add_delta(order=2)
print("Add 2-order deltas.")
else:
print("Feature type is lda+mllt")
feat = feat.splice(left=3, right=3)
feat = exkaldi.transform_feat(feat, tansformMatFile)
print("Transform feature")
## 1. Estimate the primary transform matrix from alignment or lattice.
## We estimate it from lattice, so we decode it firstly.
print("Decode the first time with original feature.")
preLat = exkaldi.decode.wfst.gmm_decode(
feat,
hmm,
HCLGfile,
wordSymbolTable=lexicons("words"),
beam=10,
latBeam=6,
acwt=0.083333,
maxActive=2000,
)
preLat.save(os.path.join(outDir, "test_premary.lat"))
print("Estimate the primary fMLLR transform matrix.")
preTransMatrix = exkaldi.hmm.estimate_fMLLR_matrix(
aliOrLat=preLat,
lexicons=lexicons,
aliHmm=hmm,
feat=feat,
adaHmm=None,
silenceWeight=0.01,
acwt=0.083333,
spk2utt=os.path.join("exp", "data", "test", "spk2utt"),
)
del preLat
## 2. Transform feature. We will use new feature to estimate the secondary transform matrix from lattice.
print("Transform feature with primary matrix.")
fmllrFeat = exkaldi.use_fmllr(
feat,
preTransMatrix,
utt2spkFile=os.path.join("exp", "data", "test", "utt2spk"),
)
print("Decode the second time with primary fmllr feature.")
secLat = exkaldi.decode.wfst.gmm_decode(
fmllrFeat,
hmm,
HCLGfile,
wordSymbolTable=lexicons("words"),
beam=13,
latBeam=6,
acwt=0.083333,
maxActive=7000,
config={"--determinize-lattice": "false"},
)
print("Determinize secondary lattice.")
thiLat = secLat.determinize(acwt=0.083333, beam=4)
print("Estimate the secondary fMLLR transform matrix.")
secTransMatrix = exkaldi.hmm.estimate_fMLLR_matrix(
aliOrLat=thiLat,
lexicons=lexicons,
aliHmm=hmm,
feat=fmllrFeat,
adaHmm=None,
silenceWeight=0.01,
acwt=0.083333,
spk2utt=os.path.join("exp", "data", "test", "spk2utt"),
)
del fmllrFeat
del thiLat
## 3. Compose the primary matrix and secondary matrix and get the final transform matrix.
print("Compose the primary and secondary transform matrix.")
finalTransMatrix = exkaldi.hmm.compose_transform_matrixs(
matA=preTransMatrix,
matB=secTransMatrix,
bIsAffine=True,
)
finalTransMatrix.save(os.path.join(outDir, "trans.ark"))
print("Transform feature with final matrix.")
## 4. Transform feature with the final transform matrix and use it to decode.
## We directly use the lattice generated in the second step. The final lattice is obtained.
finalFmllrFeat = exkaldi.use_fmllr(
feat,
finalTransMatrix,
utt2spkFile=os.path.join("exp", "data", "test", "utt2spk"),
)
del finalTransMatrix
print("Rescore secondary lattice.")
lat = secLat.am_rescore(
hmm=hmm,
feat=finalFmllrFeat,
)
print("Determinize secondary lattice.")
lat = lat.determinize(acwt=0.083333, beam=6)
lat.save(os.path.join(outDir, "test.lat"))
print("Generate lattice done.")
phoneMapFile = os.path.join("exp", "dict", "phones.48_to_39.map")
phoneMap = exkaldi.ListTable(name="48-39").load(phoneMapFile)
refText = exkaldi.load_trans(os.path.join("exp", "data", "test",
"text")).convert(phoneMap, None)
refText.save(os.path.join(outDir, "ref.txt"))
print("Generate reference text done.")
print("Now score:")
bestWER = (1000, 0, 0)
bestResult = None
for penalty in [0., 0.5, 1.0]:
for LMWT in range(1, 11):
# Add penalty
newLat = lat.add_penalty(penalty)
# Get 1-best result (word-level)
result = newLat.get_1best(lexicons("words"),
hmm,
lmwt=LMWT,
acwt=1)
# Transform from int value format to text format
result = exkaldi.hmm.transcription_from_int(
result, lexicons("words"))
# Transform 48-phones to 39-phones
result = result.convert(phoneMap, None)
# Compute WER
score = exkaldi.decode.score.wer(ref=refText,
hyp=result,
mode="present")
if score.WER < bestWER[0]:
bestResult = result
bestWER = (score.WER, penalty, LMWT)
print(f"Penalty: {penalty}, LMWT: {LMWT}, WER: {score.WER}%")
print("Score done. Save the best result.")
bestResult.save(os.path.join(outDir, "hyp.txt"))
with open(os.path.join(outDir, "best_WER"), "w") as fw:
fw.write(f"WER {bestWER[0]}, penalty {bestWER[1]}, LMWT {bestWER[2]}")
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 GMM_decode_mfcc_and_score(outDir, hmm, HCLGfile, tansformMatFile=None):
exkaldi.utils.make_dependent_dirs(outDir, pathIsFile=False)
lexicons = exkaldi.decode.graph.load_lex(
os.path.join("exp", "dict", "lexicons.lex"))
print(f"Load test feature.")
featFile = os.path.join("exp", "mfcc", "test", "mfcc_cmvn.ark")
feat = exkaldi.load_feat(featFile)
if tansformMatFile is None:
print("Feature type is delta")
feat = feat.add_delta(order=2)
print("Add 2-order deltas.")
else:
print("Feature type is lda+mllt")
feat = feat.splice(left=3, right=3)
feat = exkaldi.transform_feat(feat, tansformMatFile)
print("Transform feature")
print("Start to decode")
lat = exkaldi.decode.wfst.gmm_decode(feat,
hmm,
HCLGfile,
wordSymbolTable=lexicons("words"),
beam=13,
latBeam=6,
acwt=0.083333)
lat.save(os.path.join(outDir, "test.lat"))
print(f"Generate lattice done.")
phoneMapFile = os.path.join("exp", "dict", "phones.48_to_39.map")
phoneMap = exkaldi.ListTable(name="48-39").load(phoneMapFile)
refText = exkaldi.load_trans(os.path.join("exp", "data", "test",
"text")).convert(phoneMap, None)
refText.save(os.path.join(outDir, "ref.txt"))
print("Generate reference text done.")
print("Now score:")
bestWER = (1000, 0, 0)
bestResult = None
for penalty in [0., 0.5, 1.0]:
for LMWT in range(1, 11):
# Add penalty
newLat = lat.add_penalty(penalty)
# Get 1-best result (word-level)
result = newLat.get_1best(lexicons("words"),
hmm,
lmwt=LMWT,
acwt=1)
# Transform from int value format to text format
result = exkaldi.hmm.transcription_from_int(
result, lexicons("words"))
# Transform 48-phones to 39-phones
result = result.convert(phoneMap, None)
# Compute WER
score = exkaldi.decode.score.wer(ref=refText,
hyp=result,
mode="present")
if score.WER < bestWER[0]:
bestResult = result
bestWER = (score.WER, penalty, LMWT)
print(f"Penalty: {penalty}, LMWT: {LMWT}, WER: {score.WER}%")
print("Score done. Save the best result.")
bestResult.save(os.path.join(outDir, "hyp.txt"))
with open(os.path.join(outDir, "best_WER"), "w") as fw:
fw.write(f"WER {bestWER[0]}, penalty {bestWER[1]}, LMWT {bestWER[2]}")