def create_NMs(tacotron2_params, logger=None, decoder_infer=False):
data_preprocessor = nemo_asr.AudioPreprocessing(
**tacotron2_params["AudioPreprocessing"])
text_embedding = nemo_tts.TextEmbedding(
len(tacotron2_params["labels"]), **tacotron2_params["TextEmbedding"])
t2_enc = nemo_tts.Tacotron2Encoder(**tacotron2_params["Tacotron2Encoder"])
if decoder_infer:
t2_dec = nemo_tts.Tacotron2DecoderInfer(
**tacotron2_params["Tacotron2Decoder"])
else:
t2_dec = nemo_tts.Tacotron2Decoder(
**tacotron2_params["Tacotron2Decoder"])
t2_postnet = nemo_tts.Tacotron2Postnet(
**tacotron2_params["Tacotron2Postnet"])
t2_loss = nemo_tts.Tacotron2Loss()
makegatetarget = nemo_tts.MakeGate()
if logger:
total_weights = (text_embedding.num_weights + t2_enc.num_weights +
t2_dec.num_weights + t2_postnet.num_weights)
logger.info('================================')
logger.info(f"Total number of parameters: {total_weights}")
logger.info('================================')
return (data_preprocessor, text_embedding, t2_enc, t2_dec, t2_postnet,
t2_loss, makegatetarget)
def create_NMs(tacotron2_params, decoder_infer=False):
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
**tacotron2_params["AudioToMelSpectrogramPreprocessor"])
text_embedding = nemo_tts.TextEmbedding(
len(tacotron2_params["labels"]) + 3, # + 3 special chars
**tacotron2_params["TextEmbedding"])
t2_enc = nemo_tts.Tacotron2Encoder(**tacotron2_params["Tacotron2Encoder"])
if decoder_infer:
t2_dec = nemo_tts.Tacotron2DecoderInfer(
**tacotron2_params["Tacotron2Decoder"])
else:
t2_dec = nemo_tts.Tacotron2Decoder(
**tacotron2_params["Tacotron2Decoder"])
t2_postnet = nemo_tts.Tacotron2Postnet(
**tacotron2_params["Tacotron2Postnet"])
t2_loss = nemo_tts.Tacotron2Loss(**tacotron2_params["Tacotron2Loss"])
makegatetarget = nemo_tts.MakeGate()
total_weights = (text_embedding.num_weights + t2_enc.num_weights +
t2_dec.num_weights + t2_postnet.num_weights)
nemo.logging.info('================================')
nemo.logging.info(f"Total number of parameters: {total_weights}")
nemo.logging.info('================================')
return (data_preprocessor, text_embedding, t2_enc, t2_dec, t2_postnet,
t2_loss, makegatetarget)
def __init__(self):
"""Loads pre-trained ASR model"""
device = nemo.core.DeviceType.CPU
self.nf = nemo.core.NeuralModuleFactory(placement=device)
# Create text to spectrogram model
self.tts_conf = parse_yaml("conf.yaml")["tts"]
self.tacotron2_params = parse_yaml(
os.path.join(self.tts_conf["model_dir"], "tacotron2.yaml"))
self.nf.logger.info('================================')
# create text embedding module
self.text_embedding = nemo_tts.TextEmbedding(
len(self.tacotron2_params["labels"]) + 3, # + 3 special chars
**self.tacotron2_params["TextEmbedding"])
self.text_embedding.restore_from(
os.path.join(self.tts_conf["model_dir"], "TextEmbedding.pt"))
self.nf.logger.info(f"Number of parameters in text-embedding: "
f"{self.text_embedding.num_weights}")
# create encoder
self.t2_enc = nemo_tts.Tacotron2Encoder(
**self.tacotron2_params["Tacotron2Encoder"])
self.t2_enc.restore_from(
os.path.join(self.tts_conf["model_dir"], "Tacotron2Encoder.pt"))
self.nf.logger.info(
f"Number of parameters in encoder: {self.t2_enc.num_weights}")
# create decoder
self.t2_dec = nemo_tts.Tacotron2DecoderInfer(
**self.tacotron2_params["Tacotron2Decoder"])
self.t2_dec.restore_from(
os.path.join(self.tts_conf["model_dir"], "Tacotron2Decoder.pt"))
self.nf.logger.info(
f"Number of parameters in decoder: {self.t2_dec.num_weights}")
# create PostNet
self.t2_postnet = nemo_tts.Tacotron2Postnet(
**self.tacotron2_params["Tacotron2Postnet"])
self.t2_postnet.restore_from(
os.path.join(self.tts_conf["model_dir"], "Tacotron2Postnet.pt"))
self.nf.logger.info(
f"Number of parameters in postnet: {self.t2_postnet.num_weights}")
total_weights= self.text_embedding.num_weights+self.t2_enc.num_weights \
+ self.t2_dec.num_weights + self.t2_postnet.num_weights
self.nf.logger.info(f"Total number of parameters in model: "
f"{total_weights}")
# load waveglow if chosen
if self.tts_conf["vocoder"] == "waveglow":
self.nf.logger.info("Loading waveglow as a vocoder")
self.waveglow_params = parse_yaml(
os.path.join(self.tts_conf["vocoder_dir"], "waveglow.yaml"))
self.waveglow = nemo_tts.WaveGlowInferNM(
sigma = self.tts_conf["sigma"],
**self.waveglow_params["WaveGlowNM"])
self.waveglow.restore_from(
os.path.join(self.tts_conf["vocoder_dir"], "WaveGlowNM.pt"))
if self.tts_conf["denoising_strength"] > 0:
self.nf.logger.info("Setting up a denoiser for waveglow")
self.waveglow.setup_denoiser()
self.nf.logger.info("Waveglow denoiser is ready")
self.nf.logger.info('================================')