Audio Processing

You can find here a complete audio processing pipeline designed to prepare data for TTS applications.

You can use this code to transform a collection of audio/video material into an audio dataset suitable for training.

Pipeline structure

The pipeline follows the following structure:

• Audio track separation: extracts the audio signals and splits them into vocals and non-vocals.
• Audio Denoising: enhances and denoises the vocal signals.
• Speaker Diarization: detects different speakers thorughout the audio tracks.
• Speaker Embedding: computes speaker embeddings across all audios.
• Speaker Identification: speakers are clustered to get speakers identities across all audio data.

Requirements

Multiple environments are necessary across the pipeline.

DEMUCS

conda create -y -n demucs python=3.7
conda activate demucs
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
python3 -m pip install -U git+https://github.com/facebookresearch/demucs#egg=demucs
pip install tqdm
conda update -y ffmpeg
pip install pyannote.audio

DENOISE + PYANNOTE (SPEAKER DIARIZATION)

conda create -y -n denoiser python=3.8
conda activate denoiser
# pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install denoiser
pip install tqdm

TTS

conda create -y -n tts python=3.8
conda activate tts
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install TTS
pip install dynamicviz

How to run

In order to run the pipeline over a collection of files:

bash process_audio.sh INPUTDATAPATH DATASETNAME OUTPUTPATH

Where:

• INPUTDATAPATH: path to the data folder.
• DATASETNAME: name of the dataset to process.
• OUTPUTDATAPATH: path to the output folder.

Example:

bash process_audio.sh /home/alberto/NaturalVoices data/ /home/alberto/NaturalVoices/results

How to run on VCTK dataset

Download the dataset from: https://datashare.is.ed.ac.uk/handle/10283/2651 . Unpack the dataset and run the following:

bash process_VCTK.sh INPUTDATAPATH

Where:

• INPUTDATAPATH: path to the data folder cointaining the unpacked VCTK dataset.

Example:

bash process_VCTK.sh /mnt

Take a look at the "visualize_embeddings_VCTK.ipynb" to visualize the embeddings.

WebDataset

Here are some intructions to build a webdataset.
Webdatasets are used to easily transfer and load data. By standardizing the dataset structure, we can train multiple models with multiple dataset with less confusion and errors. \

In order to build a webdataset follow these instructions:

• Process the dataset: make a dataset specific script to build .flac - .json pairs. See examples in /datasets.
  Example Usage: python wds_process_VCTK.py
• Compress the dataset: compress the dataset using /utils/make_tar.py.
  Example Usage: python make_tar.py --input /mnt/processed_datasets/mls_german/ --output /mnt/webdataset_tar/mls_german/
• Load the dataset: test the new dataset by loading it with /utils/wds_test_tars.py.
  Example Usage: python wds_test_tars.py --tar-path /mnt/webdataset_tar/VCTK-Corpus/train/

DataLoaders

You can now load the Webdataset using specialised dataloaders.
Inherit from BaseDataloader and make your own in /dataloaders. Overwrite the process and data_tuple methods to extract the right information from your data.