def train_language_model(target, source, env):
"""Train an n-gram language model using a plain text transcript.
Uses IBM's compiled LM tools that ship with Attila. This can also be used on a segmented transcript,
in which case the n-grams are over morphs rather than words.
Sources: transcript file, n
Targets: language model file
"""
text_file = source[0].rstr()
n = source[1].read()
with temp_dir() as prefix_dir, temp_file() as vocab_file, temp_file(suffix=".txt") as sentence_file, meta_open(text_file) as text_fd:
sentences = ["<s> %s </s>" % (l) for l in text_fd]
words = set(sum([s.split() for s in sentences], []) + ["<s>", "</s>", "<UNK>"])
with meta_open(vocab_file, "w") as ofd:
ofd.write("\n".join(words))
with meta_open(sentence_file, "w") as ofd:
ofd.write("\n".join(sentences))
prefix = os.path.join(prefix_dir, "counts")
cmd = "${ATTILA_PATH}/tools/lm_64/CountNGram -n %d %s %s %s" % (n, sentence_file, vocab_file, prefix)
out, err, success = run_command(env.subst(cmd))
if not success:
return err
lm = ".".join(target[0].rstr().split(".")[0:-2])
cmd = "${ATTILA_PATH}/tools/lm_64/BuildNGram.sh -n %d -arpabo %s %s" % (n, prefix, lm)
out, err, success = run_command(env.subst(cmd), env={"SFCLMTOOLS" : env.subst("${ATTILA_PATH}/tools/lm_64")})
if not success:
return err
return None
def _run_emma(target, source, env):
with temp_file() as gold, temp_file() as guess, meta_open(source[0].rstr()) as _guess, meta_open(source[1].rstr()) as _gold:
guesses = [x for x in _guess]
words = [x.split()[0] for x in guesses]
keep = set([x for x in guesses if re.match(r"^\w+$", x) and not re.match(r".*\d.*", x)])
with meta_open(gold, "w") as gold_fd:
gold_fd.write("\n".join([x for x in _gold if x.split()[0] in keep]))
with meta_open(guess, "w") as guess_fd:
guess_fd.write("\n".join([x for x in guesses if x.split()[0] in keep]))
cmd = env.subst("python ${EMMA} -g %s -p %s -L ${LPSOLVE_PATH}" % (guess, gold), source=source, target=target)
pid = Popen(cmd.split(), stdout=PIPE)
#out, err = pid.communicate()
#prec, rec, fscore = [float(x.strip().split()[-1]) for x in out.strip().split("\n")[-3:]]
with meta_open(target[0].rstr(), "w") as ofd:
pass
#ofd.write("\t".join(["MorphP", "MorphR", "MorphF"]) + "\n")
#ofd.write("\t".join(["%.3f" % x for x in [prec, rec, fscore]]) + "\n")
return None
def run_g2p(target, source, env):
with temp_file() as tfname, meta_open(source[0].rstr()) as pl_fd:
words = set([x.split()[0].split("(")[0] for x in pl_fd])
with meta_open(tfname, "w") as t_fd:
t_fd.write("\n".join(words))
out, err, success = run_command(env.subst("%s %s/bin/g2p.py --model %s --encoding=%s --apply %s --variants-mass=%f --variants-number=%d" % (env["PYTHON"], env["OVERLAY"], source[1].rstr(), "utf-8", tfname, .9, 4)),
env={"PYTHONPATH" : env.subst("${OVERLAY}/lib/python2.7/site-packages")},
)
if not success:
return err
else:
with meta_open(target[0].rstr(), "w") as out_fd:
out_fd.write(out)
return None
def decode(target, source, env):
"""Decode some audio using a decoding network and some models (based on the example pipelines from IBM).
This is the heart, and by far the most complicated and error-prone part, of the pipeline. Basically,
the models IBM sent us are similar, but have small variations so that some need to be run
differently. This builder tries to figure out what to do based on what model files exist, and
then run the appropriate code. If it can't figure out what to run, it throws an error. It is also
aware of how many jobs the experiment has been split into, and only runs the job it was told to. Most
of the code was just slightly-adapted from the cfg.py, construct.py, and test.py files in the acoustic
models IBM sent us.
Sources: decoding network file, vocabulary file, pronunciation file, language model file
Targets: ctm transcript file, consensus network file
"""
dnet, vocabulary, pronunciations, language_model = source
out_path, tail = os.path.split(os.path.dirname(target[0].rstr()))
env.Replace(VOCABULARY_FILE=vocabulary.rstr(),
PRONUNCIATIONS_FILE=pronunciations.rstr(),
LANGUAGE_MODEL_FILE=language_model,
NETWORK_FILE=dnet,
)
cfg = CFG(env)
postThresh = 1e-04
mlpFile = env.maybe(env.subst("${MLP_FILE}"))
melFile = env.maybe(env.subst("${MEL_FILE}"))
warpFile = env.maybe(env.subst("${WARP_FILE}"))
ldaFile = env.maybe(env.subst("${LDA_FILE}"))
priorsFile = env.maybe(env.subst("${PRIORS_FILE}"))
mlp = os.path.exists(cfg.mlpFile) and "weights.mlp" in mlpFile
nmlp = os.path.exists(cfg.mlpFile) and "weights.mlp" not in mlpFile
layer = os.path.exists(env.subst("${MODEL_PATH}/layer0"))
db = dbase.DB(dirFn=dbase.getFlatDir)
fe = FeCombo(db, int(env["SAMPLING_RATE"]), env["FEATURE_TYPE"])
fe.end = fe.fmllr
fe.pcm.pcmDir = cfg.pcmDir
fe.pcm.readMode = 'speaker'
fe.norm.normMode = 1
fe.norm.normDir = env.subst("${CMS_PATH}")
fe.fmllr.fmllrDir = env.subst("${FMLLR_PATH}")
#
# from test.py
#
jid = int(env["JOB_ID"])
jnr = int(env["JOB_COUNT"])
genLat = True
genCons = True
writeLat = False
writeCons = True
cfg.useDispatcher = False
if nmlp:
chunkSize = 10
else:
chunkSize = 5
acweight = float(env.subst("${ACOUSTIC_WEIGHT}"))
db.init(cfg.dbFile, 'utterance', False, jid, jnr, chunkSize=chunkSize)
fe.mel.readFilter(melFile)
fe.mel.readWarp(warpFile)
fe.lda.readLDA(ldaFile)
se = dsearch.Decoder(speed=12, scale=acweight, lmType=32, genLat=genLat)
se.initGraph(cfg)
se.latBeam = 7
se.linkMax = 700
if mlp:
fe.ctx2 = frontend.FeCTX([fe.fmllr])
fe.ctx2.spliceN = 4
fe.ctx2.db = db
fe.mlp.depL = [fe.ctx2]
fe.mlp.db = db
fe.end = fe.mlp
fe.mlp.mlp.read(mlpFile)
fe.mlp.mlp.layerL[0].afct = Act_Rectified()
fe.mlp.mlp.layerL[1].afct = Act_Rectified()
fe.mlp.mlp.layerL[2].afct = Act_Rectified()
fe.mlp.mlp.layerL[3].afct = Act_Sigmoid()
fe.mlp.mlp.layerL[4].afct = Act_ID()
se.sc = NNScorer()
se.dnet.scorer = se.sc
se.sc.scale = acweight
se.sc.feat = fe.end.feat
se.sc.logInput = True
se.sc.readPriors(priorsFile)
pass
elif layer:
fe.ctx2 = frontend.FeCTX([fe.fmllr])
fe.ctx2.spliceN = 4
fe.ctx2.db = db
fe.end = fe.ctx2
layerL = []
for i in range(6):
l = nnet.LayerWeights()
l.name = 'layer%d'%i
l.isTrainable = False
l.initWeightFile = env.subst('${MODEL_PATH}/layer%d') % i
layerL.append(l)
if i < 5:
l = nnet.LayerSigmoid()
l.name = 'layer%d-nonl' % i
layerL.append(l)
layerL[-1].matrixOut = True
nn = nnet.NeuralNet(layerL=layerL, depL=[fe.end])
nn.db = db
nn.configure()
se.sc = NNScorer()
se.dnet.scorer = se.sc
se.sc.scale = acweight
se.sc.feat = nn.feat
se.sc.logInput = True
se.sc.readPriors(priorsFile)
elif nmlp:
se.initAM(cfg)
mlp = fe.mlp.mlp
mlp.feat = MatrixCU()
sigmoid = Act_Sigmoid()
tanh = Act_Tanh()
actid = Act_ID()
softmax = Act_Softmax()
softmax.logOutput = True
mlp.read(mlpFile)
for layerX in range(mlp.layerL.size()):
mlp.layerL[layerX].afct = sigmoid
mlp.layerL[-1].afct = actid
se.sc = NNScorer()
se.dnet.scorer = se.sc
se.sc.scale = acweight
se.sc.logInput = True
se.sc.feat = mlp.layerL[-1].Y.mat
se.sc.readPriors(priorsFile)
se.latBeam = 6.5
se.linkMax = 700
binThresh = 1.0e-10
writeSIL = 0
totUtt = 0
totArc = 0
totNonSil = 0
totDur = 0.0
totDens = 0.0
else:
return "Don't know how to run ASR with these models!"
with meta_open(target[0].rstr(), "w") as ctm_ofd, tarfile.open(target[1].rstr(), "w|gz") as tf_ofd, temp_file() as temp_fname:
for utt in db:
key = utt + ' ' + os.path.splitext(db.getFile(utt))[0]
if mlp or nmlp:
fe.end.eval(utt)
else:
nn.eval(utt)
se.search()
txt = se.getHyp().strip()
hyp = se.getCTM(key, db.getFrom(utt))
tscore = se.getScore()
for c in hyp:
ctm_ofd.write("%s\n" % (c))
se.rescore(env["RESCORE_BEAM"])
with meta_open(temp_fname, "w") as ofd:
pass
if writeLat:
fname = "%s.fsm" % (utt)
se.lat.write(temp_fname, db.getFrom(utt))
elif writeCons:
fname = "%s.cons" % (utt)
arcN = len(se.lat.arcs)
durS = db.getTo(utt)- db.getFrom(utt)
dens = arcN / durS
se.consensus(postThresh)
binThresh = 1.0e-10
writeSIL = 0
se.cons.write(temp_fname, db.getFrom(utt), binThresh, writeSIL)
tf_ofd.add(temp_fname, arcname=fname)
tf_ofd.close()
return None