Neural Formant Synthesis with Differentiable Resonant Filters

Neural formant synthesis using differtiable resonant filters and source-filter model structure.

Authors: Lauri Juvela, Pablo Pérez Zarazaga, Gustav Eje Henter, Zofia Malisz

Table of contents

1. Model overview
   1. Sound samples
2. Repository installation
   1. Conda environment
   2. GlotNet
   3. HiFi-GAN
   4. Additional libraries
3. Pre-trained models
4. Inference
5. Training
6. Citation information

Model overview

We present a model that performs neural speech synthesis using the structure of the source-filter model, allowing to independently inspect and manipulate the spectral envelope and glottal excitation:

Sound samples

A description of the presented model and sound samples compared to other synthesis/manipulation systems can be found in the project's demo webpage

Repository installation

Conda environment

First, we need to create a conda environment to install our dependencies. Use mamba to speed up the process if possible.

mamba env create -n neuralformants -f environment.yml
conda activate neuralformants

Pre-trained models are available in HuggingFace, and can be downloaded using git-lfs. If you don't have git-lfs installed (it's included in environment.yml), you can find it here. Use the following command to download the pre-trained models:

git submodule update --init --recursive

Install the package in development mode:

pip install -e .

GlotNet

GlotNet is included partially for WaveNet models and DSP functions. Full repository is available here

HiFi-GAN

HiFi-GAN is included in the hifi_gan subdirectory. Original source code is available here

Inference

We provide a script to run inference on the end-to-end architecture, such that an audio file can be provided as input and a wav file with the manipulated features is stored as output.

Change the feature scaling to modify pitch (with F0) or formants. The scales are provided as a list of 5 elements with the following order:

[F0, F1, F2, F3, F4]

An example with the provided audio samples from the VCTK dataset can be run using:

HiFi-Glot

python inference_hifiglot.py \
    --input_path "./Samples" \
    --output_path "./output/hifi-glot" \
    --config "./checkpoints/HiFi-Glot/config_hifigan.json" \
    --fm_config "./checkpoints/HiFi-Glot/config_feature_map.json" \
    --checkpoint_path "./checkpoints/HiFi-Glot" \
    --feature_scale "[1.0, 1.0, 1.0, 1.0, 1.0]"

NFS

python inference_hifigan.py \
    --input_path "./Samples" \
    --output_path "./output/nfs" \
    --config "./checkpoints/NFS/config_hifigan.json" \
    --fm_config "./checkpoints/NFS/config_feature_map.json" \
    --checkpoint_path "./checkpoints/NFS" \
    --feature_scale "[1.0, 1.0, 1.0, 1.0, 1.0]"

NFS-E2E

python inference_hifigan.py \
    --input_path "./Samples" \
    --output_path "./output/nfs-e2e" \
    --config "./checkpoints/NFS-E2E/config_hifigan.json" \
    --fm_config "./checkpoints/NFS-E2E/config_feature_map.json" \
    --checkpoint_path "./checkpoints/NFS-E2E" \
    --feature_scale "[1.0, 1.0, 1.0, 1.0, 1.0]"

Model training

Training of the HiFi-GAN and HiFi-Glot models is possible with the end-to-end architecture by using the the scripts train_e2e_hifigan.py and train_e2e_hifiglot.py.

Citation information

Citation information will be added when a pre-print is available.