VTS

What it does

API availability

The model is not packaged as a Hugging Face Inference API pipeline and is not deployed by an Inference Provider.checked 4 Oct 2026

Checkpoint

The inference code uses the pretrained checkpoint named dynamic_v3_0415.ckpt, which is available from the linked Hugging Face model repository.checked 4 Oct 2026

Checkpoint access

If the Hugging Face repository requires authentication, inference setup uses an HF_TOKEN environment variable.checked 4 Oct 2026

Deployment

The repository provides an inference-only package with local model and runtime code; training code and datasets are excluded.checked 4 Oct 2026

Generation

Generated WAV files are written to a chosen output directory, and the default output duration follows the input audio duration unless a duration is specified.checked 4 Oct 2026

Generation length

Output duration defaults to the input audio duration, can be configured, and the checkpoint is tuned for short sound-effect clips.checked 4 Oct 2026

Hardware

The quick-start inference example specifies the CUDA device.checked 4 Oct 2026

Hardware setup

The documented local requirements pin PyTorch and torchaudio CUDA 12.4 builds, with instructions to install matching builds for other CUDA drivers.checked 4 Oct 2026

How it works

It uses voice conditioning derived from dynamic audio features alongside text conditioning from a prompt.checked 4 Oct 2026

Integrations

The inference code uses google/flan-t5-base as its text encoder and includes local vocoder code.checked 4 Oct 2026

Intended users

The checkpoint is intended for research and creative sound-effect generation from vocal sketches or short audio sketches plus text prompts.checked 4 Oct 2026

License

The project and model checkpoint are listed under the MIT License.checked 4 Oct 2026

Limits

The model is optimized for short sound-effect style clips, and output quality depends on the checkpoint, input audio, prompt text, and sampling settings.checked 4 Oct 2026

Local inference

The repository provides an inference-only package with local model and runtime code.checked 4 Oct 2026

Output

Generated audio files are written as WAV files to the chosen output directory.checked 4 Oct 2026

Purpose

VTS generates sound effects from a short vocal or audio sketch combined with a text prompt.checked 4 Oct 2026

Sampling

Sampling uses a local ODE solver and typically runs 64 steps with CFG scale 3.0.checked 4 Oct 2026

Support

The repository says to contact the maker at [email protected] with questions.checked 4 Oct 2026

Text encoder

The inference path encodes text prompts with google/flan-t5-base.checked 4 Oct 2026

Training

Training code and dataset manifests are not included in the inference package.checked 4 Oct 2026

Voice conditioning

Voice conditioning uses dynamic features derived from spectral centroid, RMS, and chroma-index signals.checked 4 Oct 2026