Fast and Simple vocoder, Multiband RNN_MS.

• Demo
• Quick training
• How to Use
• System Details
• Results
• References

Demo

ToDO: Link super great impressive high-quatity audio demo.

Quick Training

Jump to ☞ , then Run. That's all!

How to Use

1. Install

# pip install "torch==1.10.0" -q      # Based on your environment (validated with v1.10)
# pip install "torchaudio==0.10.0" -q # Based on your environment
pip install git+https://github.com/tarepan/MultibandRNNMS

2. Data & Preprocessing

"Batteries Included".
RNNMS transparently download corpus and preprocess it for you 😉

3. Train

python -m mbrnnms.main_train

For arguments, check ./mbrnnms/config.py

Advanced: Other datasets

You can switch dataset with arguments.
All speechcorpusy's preset corpuses are supported.

# LJSpeech corpus
python -m mbrnnms.main_train data.data_name=LJ

Advanced: Custom dataset

Copy mbrnnms.main_train and replace DataModule.

    # datamodule = LJSpeechDataModule(batch_size, ...)
    datamodule = YourSuperCoolDataModule(batch_size, ...)
    # That's all!

System Details

Model

• PreNet: GRU
• Upsampler: time-directional nearest interpolation
• Decoder: Embedding-auto-regressive generative RNN with 10-bit μ-law encoding

Results

Output Sample

Demo

Performance

X [iter/sec] @ NVIDIA T4 on Google Colaboratory (AMP+, num_workers=8)

It takes about Ydays for full training.

References

Acknowlegements

• : Basic vocoder concept came from this paper.
• bshall/UniversalVocoding: Model and hyperparams are derived from this repository. All codes are re-written.