ethereumage / tacotron-2-persian

Tacotron 2 - Persian

Tacotron 2 - Persian

Visit this demo page to listen to some audio samples

This repository contains implementation of a Persian Tacotron model in PyTorch with a dataset preprocessor for the Common Voice dataset. For generating better quality audios, the acoustic features (mel-spectrogram) are fed to a WaveRNN model. I've included WaveRNN model in the code only for infernece purposes (no trainer included).

The source code in this repository is highly inspired by and partially copied (and also modified) from the following repostories:

Tacotron model: https://github.com/espnet/espnet
WaveRNN and some utils: https://github.com/mozilla/TTS

Model specs:

Encoder : CNN layers with batch-norm and a bi-directional lstm on top.
Decoder: 2 LSTMs for the recurrent part and a post-net on top.
Attention type: GMM v2 with k=25.

Datasets

The model is trained on audio files from one of the speakers in Common Voice Persian which can be downloaded from the link below: https://commonvoice.mozilla.org/en/datasets

Unfortunately, only a few number of speakers in the dataset have enough number of utterances for training a Tacotron model and most of the audio files have low quality and are noisy. I found audio files from one of the speakers more approriate for training whose speaker id is hard-coded in the commonvoice_fa preprocessor.

Data preprocessing

After downloading the dataset, first set DATASET_PATH and DATASET_PATH variables in the file scripts/preprocess_commonvoice_fa/preprocess_commonvoice_fa.sh and then run:

sh scripts/preprocess_commonvoice_fa/preprocess_commonvoice_fa.sh

This will extract features required for training the model and will create a meta file that contains transcripts and phonemization of each transcript in individual lines along with other meta info.
Finally, you will need to create two files named metadata_train.txt and metadata_eval.txt out of metadata.txt. First get number of lines in the transcript file with wc -l metadata.txt, then if for example there is 10000 lines in the metadata file you can split it with as below:

shuf metadata.txt >> metadata_shuf.txt
head -n 9000 metadata_shuf.txt > metadata_train.txt
tail -n 1000 metadata_shuf.txt > metadata_eval.txt

Experiments

Each experiment consists of three files namely train.sh, generate.sh and config.yml.

Training

All training parameters and model hyperparameters are saved in a YAML config file under the folder created for each experiment. To train the model for Common Voice Persian you can first change the parameters in scripts/tacotron2persian_commonvoice_fa/config.yml and then simply run:

sh scripts/tacotron2persian_commonvoice_fa/train.sh

Inference

Once the model training is finished, you can generate audio after setting the variables in generate.sh as below:

sh scripts/tacotron2persian_commonvoice_fa/generate.sh

Demo

To only test the model, I've provided a demo script folder that also contains the checkpoints for Tacotron and WaveRNN. Download the demo experiment zip from:
https://drive.google.com/file/d/1wzlBGgROS76W4xOY1UpCZaW5zropyLVL/view?usp=sharing
and unzip it under ./scripts/ folder. Similar to what explained in the Inference section, you just need to run:

sh scripts/demo/generate.sh

Example outputs

Below you see the generated mel-spectrogram and attention matrix for the setence:

صاحب قصر زنی بود با لباس‌هایِ بسیار مجلل و خدم و حشم فراوان که به گرمی از مسافرین پذیرایی کرد.

Obviously, the quality of generated mel-spectrogram/audio is lower than those generated from a standard model trained on English or Japanese languages, which is mainly caused by the quality of the audios in the dataset.

WaveRNN

The WaveRNN checkpoint used for generating the sample audios is trained on LibriTTS with 22K sampling rate. It can reconstruct waveforms for any female/male voice with an acceptable quality. Give it a try on your own voice ;)