The trained tacotron2 model is 600 MB, however in the notebook its around 300MB

Shlomo · February 7, 2021, 7:18pm

Hello,
I trained a tacotron2 model using a custom DB, by following this nice recipe:

erogol/TTS_recipes/blob/master/LJSpeech/DoubleDecoderConsistency/train_model.sh

# download LJSpeech dataset
wget http://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2
# decompress
tar -xjf LJSpeech-1.1.tar.bz2
# create train-val splits
shuf LJSpeech-1.1/metadata.csv > LJSpeech-1.1/metadata_shuf.csv
head -n 12000 LJSpeech-1.1/metadata_shuf.csv > LJSpeech-1.1/metadata_train.csv
tail -n 1100 LJSpeech-1.1/metadata_shuf.csv > LJSpeech-1.1/metadata_val.csv
# get TTS to your local
 git clone https://github.com/mozilla/TTS
# install deps
sudo apt-get install espeak
pip install soundfile
# checkout a specific version
cd TTS
git checkout b1935c97
python setup.py install
cd ..
# compute dataset mean and variance for normalization
python TTS/compute_statistics.py --config_path model_config.json --out_path ./

This file has been truncated. show original

I then wanted to test the trained model, with WaveRNN as exemplified in this notebook:
https://colab.research.google.com/github/tugstugi/dl-colab-notebooks/blob/master/notebooks/Mozilla_TTS_WaveRNN.ipynb

However, on my local system, this is the resulting size:

And in the notebook, the tacotron model is downloaded from here: https://drive.google.com/uc?id=1otOqpixEsHf7SbOZIcttv3O7pG0EadDx and its 322MB.

Naturally, my model does not load in that notebook. Can someone pinpoint the problem here?

Many thanks,

nmstoker · February 8, 2021, 12:09am

Hello @Shlomo,
Isn’t it due to different checkpoints you’re using between the two? And it’s worth comparing the config files to see what else (if anything) differs.

erogol · February 9, 2021, 3:50pm

I think what you train is Tacotron2 with Double decoder consistency and the one in the notebook is just Tacotron2. So the prior is larger than the original model. that is the difference.