def prepare_test_data(postProbDim):
feat = exkaldi.load_feat( f"{args.expDir}/train_lstm/data/test/fmllr.ark" )
if args.useCMVN:
cmvn = exkaldi.load_cmvn( f"{args.expDir}/train_lstm/data/test/cmvn_of_fmllr.ark" )
feat = exkaldi.use_cmvn(feat, cmvn, utt2spk=f"{args.expDir}/train_lstm/data/test/utt2spk")
del cmvn
if args.delta > 0:
feat = feat.add_delta(args.delta)
if args.splice > 0:
feat = feat.splice(args.splice)
feat = feat.to_numpy()
if args.normalizeFeat:
feat = feat.normalize(std=True)
# Normalize acoustic model output
if args.normalizeAMP:
ali = exkaldi.load_ali(f"{args.expDir}/train_lstm/data/train/pdfID.npy", aliType="pdfID")
normalizeBias = exkaldi.nn.compute_postprob_norm(ali,postProbDim)
else:
normalizeBias = 0
# ref transcription
trans = exkaldi.load_transcription(f"{args.expDir}/train_lstm/data/test/text")
convertTable = exkaldi.load_list_table(f"{args.expDir}/dict/phones.48_to_39.map")
trans = trans.convert(convertTable)
return feat, normalizeBias, trans
def main():
# ------------- Parse arguments from command line ----------------------
# 1. Add a discription of this program
args.discribe(
"This program is used to make dictionaries and language model")
# 2. Add options
args.add(
"--expDir",
abbr="-e",
dtype=str,
default="exp",
discription="The data resources and output path of current experiment."
)
args.add("--order",
abbr="-o",
dtype=int,
default=6,
minV=1,
maxV=6,
discription="The maximum order of N-grams language model.")
# 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", "make_dict_and_LM.args")
args.save(argsLogFile)
# ------- Make the word-pronumciation lexicon file ------
textFile = os.path.join(args.expDir, "data", "train", "text")
trainTrans = exkaldi.load_transcription(
textFile) # "trans" is an exkaldi Transcription object
wordCount = trainTrans.count_word().sort(
) # accumulate all words and their frequency in the transcription
word2pron = exkaldi.ListTable(
dict((word, word)
for word in wordCount.keys())) # word to pronunciation
pronFile = os.path.join(args.expDir, "dict", "pronunciation.txt")
word2pron.save(pronFile) # save it to file
# ------- Make lexicons ------
# 1. Generate the LexiconBank object from word-pronumciation file.
# Depending on task, about 20 lexicons will be generated and managed by the LexiconBank.
lexicons = exkaldi.decode.graph.lexicon_bank(pronFile,
silWords={"sil": "sil"},
unkSymbol={"sil": "sil"},
optionalSilPhone="sil",
extraQuestions=[],
positionDependent=False,
shareSilPdf=False,
extraDisambigPhoneNumbers=1,
extraDisambigWords=[])
# 2. Add two extra questions.
lexicons.add_extra_question(lexicons("silence_phones"))
lexicons.add_extra_question(lexicons("nonsilence_phones"))
# 3. Save this lexicon bank for future use.
lexicons.save(os.path.join(args.expDir, "dict", "lexicons.lex"))
print(f"Generate lexicon bank done.")
# ------- Make Lexicon fst ------
# 1. Generate the Lexicon fst
exkaldi.decode.graph.make_L(lexicons,
outFile=os.path.join(args.expDir, "dict",
"L.fst"),
useSilprob=0.0,
useDisambigLexicon=False)
print(f"Generate lexicon fst done.")
# 1. Generate the disambig Lexicon fst
exkaldi.decode.graph.make_L(lexicons,
outFile=os.path.join(args.expDir, "dict",
"L_disambig.fst"),
useSilprob=0.0,
useDisambigLexicon=True)
print(f"Generate disambiguation lexicon fst done.")
# ------- Make GMM-HMM topological structure for GMM-HMM ------
exkaldi.hmm.make_topology(
lexicons,
outFile=os.path.join(args.expDir, "dict", "topo"),
numNonsilStates=3,
numSilStates=5,
)
print(f"Generate topo file done.")
# ------- Train N-Grams language model ------
# 1. Train a LM.
# We have trained 2,3,4 grams model with both srilm and kenlm and chose the best one, which is 3-grams model back kenlm.
# So we directly train this one.
exkaldi.lm.train_ngrams_kenlm(
lexicons,
order=args.order,
text=
trainTrans, # If "text" received an exkaldi Transcription object, the information of utterance IDs will be omitted automatically.
outFile=os.path.join(args.expDir, "lm", f"{args.order}grams.arpa"),
config={
"--discount_fallback": True,
"-S": "20%"
},
)
print(f"Generate ARPA language model done.")
# 2. Then test this model by compute the perplexity.
exkaldi.lm.arpa_to_binary(
arpaFile=os.path.join(args.expDir, "lm", f"{args.order}grams.arpa"),
outFile=os.path.join(args.expDir, "lm", f"{args.order}grams.binary"),
)
model = exkaldi.load_ngrams(
os.path.join(args.expDir, "lm", f"{args.order}grams.binary")
) # Actually, "load_ngrams" function also accepts ARPA format file.
# 3. Prepare test transcription
testTrans = exkaldi.load_transcription(
os.path.join(args.expDir, "data", "test", "text"))
# 4. score
perScore = model.perplexity(testTrans)
print(f"The weighted average perplexity of this model is: {perScore}.")
del model
del testTrans
# ------- Make Grammar fst ------
exkaldi.decode.graph.make_G(
lexicons,
arpaFile=os.path.join(args.expDir, "lm", f"{args.order}grams.arpa"),
outFile=os.path.join(args.expDir, "lm", f"G.{args.order}.fst"),
order=args.order)
print(f"Make Grammar fst done.")
# ------- Compose LG fst for futher use ------
exkaldi.decode.graph.compose_LG(
LFile=os.path.join(args.expDir, "dict", "L_disambig.fst"),
GFile=os.path.join(args.expDir, "lm", f"G.{args.order}.fst"),
outFile=os.path.join(args.expDir, "lm", f"LG.{args.order}.fst"),
)
print(f"Compose LG fst done.")
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 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 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"),
)
def main():
# 1, Parse command line options.
args.add("--order",
abbr="-o",
dtype=int,
default=[2, 3, 4, 5, 6],
minV=1,
maxV=6,
discription="The language model order.")
args.add("--dataDir",
abbr="-d",
dtype=str,
default="./exp",
discription="The resource directory.")
args.add("--exp",
abbr="-e",
dtype=str,
default="./exp_lms",
discription="Experiment output directory.")
args.parse()
# 2, Prepare text and lexicon.
declare.is_file(
f"./{args.dataDir}/data/train/text",
debug=
"There is not train text file avaliable. Please run '01_prepare_data.py' to generate it."
)
declare.is_file(
f"./{args.dataDir}/data/test/text",
debug=
"There is not train text file avaliable. Please run '01_prepare_data.py' to generate it."
)
declare.is_file(
f"./{args.dataDir}/dict/lexicons.lex",
debug=
"There is not lexicon file avaliable. Please run '02_make_dict_and_LM.py' to generate it."
)
trainTrans = exkaldi.load_transcription(
f"./{args.dataDir}/data/train/text")
lexicons = exkaldi.load_lex(f"./{args.dataDir}/dict/lexicons.lex")
testTrans = exkaldi.load_transcription(f"./{args.dataDir}/data/test/text")
# 3, Train LMs and compute perplexities.
for o in args.order:
for backend in ["sri", "ken"]:
if backend == "sri":
exkaldi.lm.train_ngrams_srilm(
lexicons,
order=o,
text=
trainTrans, # If "text" received an exkaldi Transcription object, the information of utterance IDs will be omitted automatically.
outFile=os.path.join(args.exp, f"{backend}_{o}grams.arpa"),
config={"-wbdiscount": True},
)
else:
exkaldi.lm.train_ngrams_kenlm(
lexicons,
order=o,
text=
trainTrans, # If "text" received an exkaldi Transcription object, the information of utterance IDs will be omitted automatically.
outFile=os.path.join(args.exp, f"{backend}_{o}grams.arpa"),
config={
"--discount_fallback": True,
"-S": "20%"
},
)
exkaldi.lm.arpa_to_binary(
arpaFile=os.path.join(args.exp, f"{backend}_{o}grams.arpa"),
outFile=os.path.join(args.exp, f"{backend}_{o}grams.binary"),
)
model = exkaldi.load_ngrams(
os.path.join(args.exp, f"{backend}_{o}grams.binary"))
perScore = model.perplexity(testTrans)
print(f"{o} {backend} score:", perScore)
