Practical overview
When it fits
Use it when you want code and preprocessing tools for a PyTorch text-to-speech experiment, especially when working with the listed datasets or adapting a documented multi-speaker workflow.
Who it is for
- Developers exploring or building neural text-to-speech systems in Python and PyTorch.
- Researchers experimenting with single-speaker or multi-speaker speech synthesis.
- Users preparing supported speech datasets for model training.
Documented uses
- Preprocess LJSpeech, VCTK, JSUT or NIKL data into features for speech-model workflows.
- Train single- or multi-speaker text-to-speech models with repository presets.
- Generate speech from text with a checkpoint using the synthesis scripts.
- Explore sample audio and single- or multi-speaker demo notebooks.
Requirements & limits
The repository documents local installation and Colab demo notebooks. Japanese preprocessing requires MeCab.
- The documented local environment requires Python 3.5+, CUDA 8.0+, PyTorch 1.0+ and nnmnkwii 0.0.11+.
- MeCab is an additional requirement for Japanese.
- The repository warns that listed pretrained checkpoints are incompatible with the current master branch and recommends using their specified Git commits.
- Preset parameters should remain consistent across preprocessing, training and evaluation.
- VCTK multi-speaker preparation includes phoneme alignment and recommended silence removal.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Preprocess LJSpeech with the DeepVoice3 preset
Prepare LJSpeech audio and text data into model features using the repository’s documented preset workflow.
Before you start
- Python 3.5 or newer, CUDA 8.0 or newer, PyTorch 1.0 or newer, and nnmnkwii 0.0.11 or newer.
- A local LJSpeech dataset.
- The repository cloned locally and its listed dependencies installed.
Steps
Clone the repository and install it with the command below.
pip install -e ".[bin]"Run the command below.
python preprocess.py --preset=presets/deepvoice3_ljspeech.json ljspeech ~/data/LJSpeech-1.0/ ./data/ljspeech
Expected outcome
The README says preprocessing produces mel-spectrogram and linear-spectrogram features in `./data/ljspeech`.
Guide limitations
- The guide covers preprocessing only, not training or synthesis.
- Use the same preset for subsequent training and evaluation stages.
Prepare VCTK data for multi-speaker training
Extract features from VCTK recordings and start the documented multi-speaker training workflow.
Before you start
- Python 3.5 or newer, CUDA 8.0 or newer, PyTorch 1.0 or newer, and nnmnkwii 0.0.11 or newer.
- A local VCTK dataset and phoneme alignments for each utterance.
- The repository’s VCTK preprocessing instructions for handling long silences.
Steps
Prepare phoneme alignments and remove silences as recommended for VCTK recordings.
Run the command below.
python preprocess.py vctk ${your_vctk_root_path} ./data/vctkRun the documented
train.pycommand with--data-root=./data/vctk,--checkpoint-dir=checkpoints_vctk, and--preset=presets/deepvoice3_vctk.json.
Expected outcome
The workflow extracts features and begins training a multi-speaker DeepVoice3 model using the VCTK preset.
Guide limitations
- The README recommends alignment and silence removal because long silences can affect performance.
- The instructions do not establish the outcome or quality of a completed training run.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- r9y9/deepvoice3_pytorch · Official homepage title · retained in its source language