FastSpeech 2 checkpoints for English and Telugu
Checkpoints for two separately trained FastSpeech 2 models, built on ming024/FastSpeech2. Code, results and the technical report: https://github.com/23f2001033/IITH-Intern-Assignment-
| File | Model | Data | Step |
|---|---|---|---|
english_step60000.pth.tar |
Multi-speaker English, with optimizer state | Svarah, about 5 h train / 1 h validation | 60,000 |
telugu_step13000.pth.tar |
Single-speaker Telugu, model weights only | IndicTTS Telugu, about 5 h train / 1 h validation | 13,000 (lowest validation loss) |
Both use the universal HiFi-GAN vocoder at 22,050 Hz, which is included in the GitHub repository.
Usage
Clone the GitHub repository, put the files in English/checkpoints/ and Telugu/checkpoints/, then:
pip install -r requirements.txt
python English/inference.py --ckpt_file English/checkpoints/english_step60000.pth.tar --results_dir out/English
python Telugu/inference.py --ckpt_file Telugu/checkpoints/telugu_step13000.pth.tar --results_dir out/Telugu
The default English voice is spk_f0bb6d4e; use --speaker to choose another, and --text to synthesise
your own sentence.