Help with extending functionality of synth_wav.sh?

I’m working on extending TTS to the languages and speakers available in /egs/m_ailabs but synth_wav.sh will only output a default speaker.

I have followed this procedure: executed run.sh through step 2, brought in a tacotron checkpoint, executed one epoch of training to obtain a proper snapshot, executed run.sh Stage 4 Network Decoding, and then the following for wav synthesis.

../../../utils/synth_wav.sh --models ljspeech.tacotron2.v3 \
	--stage 3 \
	--stop_stage 4 \
	--cmvn data/fr_FR_bernard_train_trim/cmvn.ark \
	--dict data/lang_1char/fr_FR_bernard_train_trim_units.txt \
	--synth_model exp/fr_FR_bernard_train_trim_pytorch_train_pytorch_tacotron2.v3/results/model.last1.avg.best \
	--decode_config conf/decode.yaml \
	--vocoder_models ljspeech.wavenet.softmax.ns.v1 \
	example.txt

How are we able to produce synthesized voice in another language and voice? Also, can direct steps be taken to modify the produced voice further? Thank you for your guidance and I look forward to contributing to your community.

[kan-bayashi]

It seems this issue is solved. I will close this issue.

[adhamel]

@b-flo thank you for sharing this. I’m sure it will steer me in the correct direction.

@kan-bayashi I really appreciate all of your help here. Parallel WaveGAN is converging very nicely. The discriminator loss (started training at 100k) is the only stubborn decline, hovering around ~.496-.497. If you like, after training on this different data set. I can let you know if I find any of the problems with strange noise you mention.