smtiitm/Fastspeech2_HS

Indic TTS for Indian Languages: This is a project on developing text-to-speech (TTS) synthesis systems for Indian languages, improving quality of synthesis, as well as small foot print TTS integrated with disability aids and various other applications.

Fastspeech2 Model using Hybrid Segmentation (HS)

This repository contains a Fastspeech2 Model for 16 Indian languages (male and female both) implemented using the Hybrid Segmentation (HS) for speech synthesis. The model is capable of generating mel-spectrograms from text inputs and can be used to synthesize speech. The Architecture of FS2: As requested here is the details of FS2 architecture:

Fs2 is composed of 6 feed-forward Transformer blocks with multi-head self-attention and 1D convolution on both phoneme encoder and mel-spectrogram decoder. In each feed-forward Transformer, the hidden size of multi-head attention is set to 256 and the number of head is set to 2. The kernel size of 1D convolution in the two-layer convolution network is set to 9 and 1, and the input/output size of the number of channels in the first and the second layer is 256/1024 and 1024/256. The duration predictor and variance adaptor, which are composed of stacks of several convolution networks and the final linear projection layer. The convolution layers of the duration predictor and variance adaptor are set to 2 and 5, the kernel size is set to 3, the input/output size of all layers is 256/256, and the dropout rate is set to 0.5.

The Repo is large in size. Directly go to the next Installation part. We have used Git LFS due to Github's size constraint (please install latest git LFS from the link, we have provided the current one below).

curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.python.sh | bash
sudo apt-get install git-lfs
git lfs install

Language model files are uploaded using git LFS. so please use:

git lfs fetch --all
git lfs pull

to get the original files in your directory.

Model Files

The model for each language includes the following files:

• config.yaml: Configuration file for the Fastspeech2 Model.
• energy_stats.npz: Energy statistics for normalization during synthesis.
• feats_stats.npz: Features statistics for normalization during synthesis.
• feats_type: Features type information.
• pitch_stats.npz: Pitch statistics for normalization during synthesis.
• model.pth: Pre-trained Fastspeech2 model weights.

Installation

1. Install Miniconda first. Create a conda environment using the provided environment.yml file:

conda env create -f environment.yml

2.Activate the conda environment (check inside environment.yaml file):

conda activate tts-hs-hifigan

3. Install PyTorch separately (you can install the specific version based on your requirements):

conda install pytorch cudatoolkit
pip install torchaudio

Vocoder

For generating WAV files from mel-spectrograms, you can use a vocoder of your choice. One popular option is the HIFIGAN vocoder (Clone this repo and put it in the current working directory). Please refer to the documentation of the vocoder you choose for installation and usage instructions.

(We have used the HIFIGAN vocoder and have provided Vocoder tuned on Aryan and Dravidian languages)

Usage

The directory paths are Relative. ( But if needed, Make changes to text_preprocess_for_inference.py and inference.py file, Update folder/file paths wherever required.)

Please give language/gender in small cases and sample text between quotes. Adjust output speed using the alpha parameter (higher for slow voiced output and vice versa). Output argument is optional; the provide name will be used for the output file.

Use the inference file to synthesize speech from text inputs:

python inference.py --sample_text "Your input text here" --language <language> --gender <gender> --alpha <alpha> --output_file <file_name.wav OR path/to/file_name.wav>

Example:

python inference.py --sample_text "श्रीलंका और पाकिस्तान में खेला जा रहा एशिया कप अब तक का सबसे विवादित टूर्नामेंट होता जा रहा है।" --language hindi --gender male --alpha 1 --output_file male_hindi_output.wav

The file will be stored as male_hindi_output.wav and will be inside current working directory. If --output_file argument is not given it will be stored as <language>_<gender>_output.wav in the current working directory.

Citation

If you use this Fastspeech2 Model in your research or work, please consider citing:

“ COPYRIGHT 2023, Speech Technology Consortium,

Bhashini, MeiTY and by Hema A Murthy & S Umesh,

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING and ELECTRICAL ENGINEERING, IIT MADRAS. ALL RIGHTS RESERVED "

Shield:

This work is licensed under a Creative Commons Attribution 4.0 International License.