Failed to build custom Tacotron2 in jarvis-speech-maker

AI & Data Science Deep Learning (Training & Inference) Riva
1

I trained Tacotron2 in Thai language using NeMo and it failed to build jmir format, but English Tacotron2 downloaded from NGC is build successfully.

Screen Shot 2564-05-05 at 13.58.19
Screen Shot 2564-05-05 at 13.58.191898×704 491 KB

2

Screen Shot 2564-05-05 at 14.19.23
Screen Shot 2564-05-05 at 14.19.231506×468 210 KB

3

Hi @6370120621
Could you please share the model file so we can help better?
As per error msg it seems script is not able to find encoder-model_weights.ckpt. Could you please verify the trained model has all the required files?

Thanks

4

Hi @SunilJB
This is my nemo model file Tacotron2.nemo - Google Drive

I also tried to use nemo file from TTS En Tacotron2 | NVIDIA NGC, it errors the same as my file, but use file from Speech Synthesis English Tacotron2 | NVIDIA NGC, it’s works.

What is the difference between these two? or I miss something to convert nemo file that I got after finish training to deployable?

5

Now, I fixed this issue by following this Failed to build Tacotron2 in Jarvis · Issue #2154 · NVIDIA/NeMo (github.com), but it has an error when I use command deploy_all_model in jarvis-speech-servicemaker version 1.0b3.

Screen Shot 2564-05-12 at 14.53.39
Screen Shot 2564-05-12 at 14.53.391614×300 220 KB

I use waveglow from Speech Synthesis Waveglow | NVIDIA NGC.