Practical overview
When it fits
Use Voicerestore when a speech-focused model is needed to clean up degraded voice recordings locally or as part of a Python workflow, including longer files handled through window chunking.
Who it is for
- Developers building speech-cleanup or audio preprocessing workflows
- Researchers exploring speech restoration models
- People handling degraded voice recordings who can use a Python-based model
Documented uses
- Reduce noise and reverberation in speech recordings
- Improve speech affected by distortion or signal loss
- Preprocess degraded speech before transcription
- Process longer recordings using window-chunked inference
Requirements & limits
The repository is public and MIT-licensed. Its documented local workflow requires installing dependencies and downloading a pretrained checkpoint.
- The model is a 300M-plus-parameter flow-matching transformer, so local use involves downloading a checkpoint and installing project dependencies.
- The README provides separate inference scripts for short and long recordings; the long-form script uses window chunking.
- The model is optimized for speech and may be less suitable for music or other audio.
- The repository credits the E2-TTS implementation by Lucidrains as its basis.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up Voicerestore and run a short-audio restoration
Prepare the repository and use its documented short-inference command to save a restored speech file.
Before you start
- Python environment capable of installing the repository requirements
- Git
- A degraded speech input file such as test_input.wav
- A downloaded pretrained checkpoint placed in the repository's checkpoints folder
Steps
Clone the repository with submodules using the README's git clone command, or initialize submodules after cloning.
Install dependencies with pip install -r requirements.txt.
Download the pretrained model checkpoint and place it in the checkpoints folder.
Run inference_short.py with the checkpoint, input, output, steps and cfg_strength arguments shown in the README.
Expected outcome
The documented command saves the restored result as test_output.wav.
Guide limitations
- The model is optimized for speech and may not perform optimally on music or other audio.
- The README says performance on extreme degradations remains an area of ongoing research.
Restore a long speech recording with window chunking
Use the repository's long-form inference workflow to process a longer audio file.
Before you start
- A cloned repository with dependencies installed
- A pretrained checkpoint in the checkpoints folder
- A long speech audio file such as long_audio_file.mp3
Steps
Follow the repository setup steps to install dependencies and place the checkpoint in checkpoints.
Run inference_long.py with the checkpoint, input and output paths, steps, cfg_strength, window_size_sec and overlap arguments shown in the README.
Expected outcome
The documented workflow saves the processed audio as test_output_long.wav and uses window chunking.
Guide limitations
- The model is optimized for speech; the documentation cautions that music and other audio may not perform as well.
- The README notes ongoing work on extreme degradation performance.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- skirdey/voicerestore · Official homepage title · retained in its source language