How to Train FastPitch with custom labels?

Accelerated Computing Intelligent Video Analytics TAO Toolkit
#1

Please provide the following information when requesting support.

• Hardware : NVIDIA TESLA V100
• Network Type : FastPitch
• Training spec file : FastPitch train specs (yaml)
• How to reproduce the issue ? : change notations in train.yaml into chars

Hello. I want to train Text to Speech FastPitch model using custom labels. In Training spec file, it used phonemes as default value for notations. I tried to change the value into chars, but it produce error as below :

TypeError: Error instantiating ‘nemo.collections.asr.data.audio_to_text.AudioToCharWithPriorAndPitchDataset’ : init() missing 1 required positional argument: 'labels’

How to add those argument ? I already tried to add labels argument in train.yaml file and add +labels argument in running command, neither of them work. Still give me the same error.

Thank you.

#2

May I know if you meet the same issue with official released jupyter notebook?
TAO Toolkit Quick Start Guide — TAO Toolkit 3.21.11 documentation →
Text to Speech Notebook | NVIDIA NGC → GPU-optimized AI, Machine Learning, & HPC Software | NVIDIA NGC

#3

Hello. Thank you for the answer.

The tutorial in official released jupyter notebook is working just fine. But I need to use chars type in the vocab to exactly match my sentences in the metadata file. is that possible to do that?

Thank you

#4

Please double check the dataset.
More info can be found in NeMo/audio_to_text.py at main · NVIDIA/NeMo · GitHub and NeMo/vocabs.py at 213d6685d8adfb943ba763d1c7e1e4eb9c68fb62 · NVIDIA/NeMo · GitHub

You can debug below files directly inside tao docker.
/opt/conda/lib/python3.8/site-packages/nemo/collections/asr/data/audio_to_text.py
/opt/conda/lib/python3.8/site-packages/nemo/collections/asr/data/vocabs.py

#5

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.