vocos

What it is and what it does

Vocos is a neural vocoder—a machine learning model that converts acoustic feature representations into audio waveforms. Unlike traditional vocoders that work in the time domain, Vocos generates spectral coefficients and reconstructs audio via inverse Fourier transform, enabling fast single-pass synthesis. It is trained using a GAN objective and can accept either mel-spectrograms or EnCodec tokens as input, making it suitable for integration into text-to-speech pipelines or audio processing workflows.

The package includes pre-trained models for 24 kHz audio synthesis and supports both inference and training modes. It depends on torch, torchaudio, numpy, scipy, einops, pyyaml, huggingface-hub, and encodec. Installation is straightforward, though the dormant maintenance status (last release 2023-10-14) means bug fixes and feature updates are not actively rolling out.

Use it for:

• Convert mel-spectrograms from a text-to-speech model into high-quality audio waveforms for end-to-end TTS synthesis.
• Reconstruct audio from EnCodec-compressed tokens at various bandwidth levels for codec-based audio processing.
• Perform copy-synthesis by resampling an audio file to 24 kHz and reconstructing it through the vocoder.
• Integrate with text-to-audio models as a replacement vocoder for faster or higher-quality audio generation.
• Train a custom vocoder on domain-specific audio data using the provided training pipeline and configuration framework.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Vocos is a neural vocoder that synthesizes audio waveforms from acoustic features (mel-spectrograms or EnCodec tokens) using a GAN-based model that generates spectral coefficients for fast reconstruction via inverse Fourier transform.

Yes, if you need a fast neural vocoder for mel-spectrogram or EnCodec token-to-audio synthesis and can work with a dormant codebase. Low install friction and zero known vulnerabilities make it practical for inference. Verify the license status directly in the repository before commercial use, and be aware that maintenance has stalled since 2023-10-14—expect no active support or updates.

Install

vocos on PyPI

pip install vocos

uv add vocos

poetry add vocos

Installing vocos

Before you install

Low install friction with a pure-Python wheel. Maintenance is dormant—last release was 2023-10-14 and last commit 2024-08-07—but the repository remains unarchived with moderate popularity (1150 stars).

License in practice

License treatment is unclear; the repository states MIT in its LICENSE file, but the PyPI metadata does not declare it formally. Verify the LICENSE file directly before relying on the package in a commercial or license-sensitive context.

Quickstart

pip install vocos

import torch
from vocos import Vocos

vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
mel = torch.randn(1, 100, 256)  # B, C, T
audio = vocos.decode(mel)

Requires torch and torchaudio; pre-trained models are downloaded from huggingface-hub on first use.

Verify before relying

• Whether the MIT license statement in the repository's LICENSE file is the authoritative license for the PyPI package.
• Whether the package is actively maintained or if dormancy signals a shift to a successor or fork.
• Compatibility with modern PyTorch and torchaudio versions beyond what the fact sheet specifies.

Package facts

Evidence: vocos-0.1.0-py3-none-any.whl

Tags

More Artificial Intelligence packages