def train(opts, dirs):
## Handle corpus:
print(" -- Gather corpus")
## Get names of directories containing corpus data (all txt and wav):
corpora = []
if opts.command_line_corpus:
for location in opts.command_line_corpus:
assert os.path.isdir(location)
corpora.append(location)
else:
corpora.append(
os.path.join(dirs['CORPUS'], opts.lang, fname.SPEAKERS,
opts.speaker, "txt"))
corpora.append(
os.path.join(dirs['CORPUS'], opts.lang, fname.SPEAKERS,
opts.speaker, "wav"))
# additional large text corpus:
if opts.text_corpus_name:
corpora.append(
os.path.join(dirs['CORPUS'], opts.lang, fname.TEXT_CORPORA,
opts.text_corpus_name))
# Set file number
if opts.file_num:
file_num = int(opts.file_num)
else:
file_num = float("inf")
# Get names of individual txt and wav files:
voice_data = []
for c in corpora:
count = 0
file_list = sorted(os.listdir(c))
if opts.shuffle:
random.seed(1)
random.shuffle(file_list)
for f in file_list:
if '._' not in f:
voice_data.append(os.path.join(c, f))
count += 1
# Stop appending voice data at file_num
if count >= file_num:
break
corpus = Corpus.Corpus(voice_data)
print(" -- Train voice")
voice = Voice(opts.speaker, opts.lang, opts.config, opts.stage, \
dirs, clear_old_data=opts.clear, max_cores=opts.max_cores)
## Train the voice (i.e. train processors in pipeline context):
voice.train(corpus)
def train(opts, dirs):
## Handle corpus:
print " -- Gather corpus"
## Get names of directories containing corpus data (all txt and wav):
corpora = []
if opts.command_line_corpus:
for location in opts.command_line_corpus:
assert os.path.isdir(location)
corpora.append(location)
else:
corpora.append(
os.path.join(dirs['CORPUS'], opts.lang, fname.SPEAKERS,
opts.speaker, "txt"))
corpora.append(
os.path.join(dirs['CORPUS'], opts.lang, fname.SPEAKERS,
opts.speaker, "wav"))
# additional large text corpus:
if opts.text_corpus_name:
corpora.append(
os.path.join(dirs['CORPUS'], opts.lang, fname.TEXT_CORPORA,
opts.text_corpus_name))
## Get names of individual txt and wav files:
voice_data = []
for c in corpora:
for f in os.listdir(c):
voice_data.append(os.path.join(c, f))
corpus = Corpus.Corpus(voice_data)
print " -- Train voice"
voice = Voice(opts.speaker, opts.lang, opts.config, opts.stage, \
dirs, clear_old_data=opts.clear, max_cores=opts.max_cores)
## Train the voice (i.e. train processors in pipeline context):
voice.train(corpus)
def main_work():
#################################################
# root is one level below this file in directory structure, ie. below the 'scripts' folder
ROOT = os.path.split(
os.path.realpath(
os.path.abspath(
os.path.dirname(inspect.getfile(
inspect.currentframe())))))[0] + '/'
dirs = {
'ROOT': ROOT,
'CONFIG': ROOT + "configs/",
'VOICES': ROOT + "voices/",
'TRAIN': ROOT + "train/",
'RULES': ROOT + "rules/",
'CORPUS': ROOT + "corpus/",
'BIN': ROOT + "/tools/bin/"
}
# ======== Get stuff from command line ==========
a = ArgumentParser()
a.add_argument('-s', dest='speaker', required=True, \
help= "the name of the speaker: <ROOT>/corpus/<LANG>/<SPEAKER>")
a.add_argument('-l', dest='lang', required=True, \
help= "the language of the speaker: <ROOT>/corpus/<LANG>")
a.add_argument('-o', dest='output', required=False, default=False, \
help= "output audio here")
a.add_argument('-t', dest='stage', required=False, default="runtime", \
help=""" defines the current usage stage
(definitions of stages should by found in <config>/recipe.cfg""")
a.add_argument('-play', dest='play', action="store_true", required=False, default=False, \
help=" play audio after synthesis")
a.add_argument('-lab', dest='make_label', action="store_true", default=False, \
help= "make label file as well as wave in output location")
a.add_argument('config',
help="""configuration to use: naive, semi-naive, gold,
as defined in <ROOT>/recipes/<config> -directory"""
)
a.add_argument('-bin', dest='custom_bindir')
a.add_argument('files',
nargs='*',
help="text files to speak, reading from stdin by default")
a.add_argument('-m',
dest='model_dir',
required=True,
type=str,
help="model directory")
opts = a.parse_args()
dirs['TRAIN'] = opts.model_dir + "/train/"
dirs['VOICES'] = opts.model_dir + "/voices/"
if opts.custom_bindir != None:
dirs['BIN'] = opts.custom_bindir
voice_location = os.path.join(dirs['VOICES'], opts.lang, opts.speaker,
opts.config)
train_location = os.path.join(dirs['TRAIN'], opts.lang, "speakers",
opts.speaker, opts.config)
config_path = os.path.join(dirs['CONFIG'], opts.config)
voice_config = os.path.join(config_path, fname.RECIPE)
## Make Voice object to contain voice elements trained on this corpus:
voice = Voice(opts.speaker, opts.lang, opts.config, opts.stage, dirs)
if not opts.output:
output_wavefile = os.path.join(voice_location, 'output', 'wav',
'temp.wav')
else:
output_wavefile = opts.output
if not opts.output:
output_labfile = None
else:
output_labfile = output_wavefile.replace('.wav', '.lab')
prevspace = False
para = []
# Go through the files a paragraph at a time, unless it's SSML in which case we parse it
# An empty line marks the change of paragraphs in plain text files
for line in fileinput.input(opts.files):
line = line.decode('utf-8').rstrip()
t = start_clock('Synthesise sentence')
print line
if fileinput.isfirstline():
if para != []:
voice.synth_utterance(''.join(para), output_wavefile=output_wavefile, \
output_labfile=output_labfile)
if opts.play:
os.system('play ' + output_wavefile)
para = []
line = line.lstrip()
if line.startswith('<speak') or line.startswith('<xml'):
tree = etree.parse(fileinput.filename())
parseSSML(tree, voice)
fileinput.nextfile()
else:
para.append(line)
elif line.isspace():
prevspace = True
elif prevspace and para != []:
voice.synth_utterance(''.join(para), output_wavefile=output_wavefile, \
output_labfile=output_labfile)
prevspace = False
para = [line]
else:
para.append(line)
if para != []:
voice.synth_utterance(''.join(para), output_wavefile=output_wavefile, \
output_labfile=output_labfile)
if opts.play:
os.system('play ' + output_wavefile)
stop_clock(t)
def main_work():
#################################################
# ======== Get stuff from command line ==========
def usage():
print("Usage: ...... ")
sys.exit(1)
# e.g.
try:
voice_config = sys.argv[1]
voice_components = sys.argv[2]
ENGINE_BIN = sys.argv[3]
RESYNTH_BIN = sys.argv[4]
trained_model_dir = sys.argv[5]
except:
usage()
#################################################
sys.path.append("/afs/inf.ed.ac.uk/user/o/owatts/naive/script/")
#################################################
## Lots of these paths should be interpolated from system-wide options (e.g. bin dir etc).
## Absolute paths for now.
context_file_location = "/afs/inf.ed.ac.uk/user/o/owatts/naive/context_files/"
ESTDIR = "/group/project/nlp-speech/bin/"
HTSDIR = "/afs/inf.ed.ac.uk/user/o/owatts/repos/simple4all/CSTRVoiceClone/trunk/bin/"
SCRIPT = "/afs/inf.ed.ac.uk/user/o/owatts/naive/script"
GENSIM_LOCATION = "%s/gensim-0.5.0/src/" % (SCRIPT)
#################################################
sys.path.append(GENSIM_LOCATION) ## add gensim to path
from VSMTagger import VSMTagger
print(" -- Open the existing voice")
voice = Voice(config_file=voice_config)
print(" -- Make an utterance processor from a (trained) acoustic model ")
### This will only perform work where an utt does not have a wavefile attached:
parameter_generator = AcousticModel(config_file=voice_components +
"/parameter_generator.cfg",
processor_name="parameter_generator",
ENGINE_BIN=ENGINE_BIN,
model_location=trained_model_dir,
HTSDIR=HTSDIR)
parameter_generator.save()
### WAVESYNTH
waveform_synthesiser = WaveSynthesiser(
config_file=voice_components + "/waveform_synthesiser.cfg",
processor_name="waveform_synthesiser",
RESYNTH_BIN=RESYNTH_BIN,
HTSDIR=HTSDIR)
waveform_synthesiser.save()
### WAVE PLAYER (call e.g. sox etc)
wave_player = WavePlayer(config_file=voice_components + "/wave_player.cfg",
processor_name="wave_player")
wave_player.save()
voice.add_processor(voice_components + "/parameter_generator.cfg")
voice.add_processor(voice_components + "/waveform_synthesiser.cfg")
voice.add_processor(voice_components + "/wave_player.cfg")
print(" -- Save voice")
voice.save()
print(" -- Synthesize a test utterance (from some Spanish text...)")
## Use the voice to synth a test utterance:
voice.synth_utterance("Esto es: una prueba.")
def main_work():
#################################################
# root is one level below this file in directory structure, ie. below the 'scripts' folder
ROOT = os.path.split(
os.path.realpath(
os.path.abspath(
os.path.dirname(inspect.getfile(
inspect.currentframe())))))[0] + '/'
dirs = {
'ROOT': ROOT,
'CONFIG': ROOT + "configs/",
'VOICES': ROOT + "voices/",
'TRAIN': ROOT + "train/",
'RULES': ROOT + "rules/",
'CORPUS': ROOT + "corpus/",
'BIN': ROOT + "/tools/bin/"
}
# ======== Get stuff from command line ==========
a = ArgumentParser()
a.add_argument('-s', dest='speaker', required=True, \
help= "the name of the speaker: <ROOT>/corpus/<LANG>/<SPEAKER>")
a.add_argument('-l', dest='lang', required=True, \
help= "the language of the speaker: <ROOT>/corpus/<LANG>")
a.add_argument('-o', dest='output', required=False, default=False, \
help= "output audio here")
a.add_argument('-t', dest='stage', required=False, default="runtime", \
help=""" defines the current usage stage
(definitions of stages should by found in <config>/recipe.cfg""")
a.add_argument('-u', dest='output_utt', required=False, default=False, \
help= "output utt files here")
a.add_argument('-play', dest='play', action="store_true", required=False, default=False, \
help=" play audio after synthesis")
a.add_argument('-lab', dest='make_label', action="store_true", default=False, \
help= "make label file as well as wave in output location")
a.add_argument('config',
help="""configuration to use: naive, semi-naive, gold,
as defined in <ROOT>/recipes/<config> -directory"""
)
a.add_argument('files',
nargs='*',
help="text files to speak, reading from stdin by default")
opts = a.parse_args()
voice_location = os.path.join(dirs['VOICES'], opts.lang, opts.speaker,
opts.config)
train_location = os.path.join(dirs['TRAIN'], opts.lang, "speakers",
opts.speaker, opts.config)
config_path = os.path.join(dirs['CONFIG'], opts.config)
voice_config = os.path.join(config_path, fname.RECIPE)
## Make Voice object to contain voice elements trained on this corpus:
voice = Voice(opts.speaker, opts.lang, opts.config, opts.stage, dirs)
if not opts.output:
output_dir = os.path.join(voice_location, 'output', 'wav')
else:
output_dir = opts.output
# if not opts.output:
# output_labfile = None
# else:
# output_labfile = output_wavefile.replace('.wav', '.lab')
if opts.output_utt:
if not os.path.isdir(opts.output_utt):
os.makedirs(opts.output_utt)
output_extensions = []
if opts.make_label:
output_extensions.append('dnn_lab')
for filename in opts.files:
base = get_basename(filename)
output_wavefile = os.path.join(output_dir, base + '.wav')
text = ' '.join(readlist(filename))
try:
print(text)
except:
print(' ') ## weird characgers
print(base)
if opts.output_utt:
voice.synth_utterance(text, output_wavefile=output_wavefile, \
output_uttfile=os.path.join(opts.output_utt, base + '.utt'),\
output_extensions=output_extensions)
else:
voice.synth_utterance(text, output_wavefile=output_wavefile, \
output_extensions=output_extensions)