gemelo-ai/vocos

Health v2 · maintenance only

Adoption not part of the score

1155 stars · 134 forks observed · 2026-09-01

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Vocos is a fast neural vocoder that synthesizes audio waveforms from acoustic features such as mel-spectrograms or EnCodec tokens. It uses a GAN trained in the frequency domain, generating spectral coefficients and reconstructing audio via inverse Fourier transform for an order-of-magnitude speedup over time-domain vocoders.

Use cases

• convert mel-spectrograms to audio waveforms
• decode EnCodec tokens back to audio
• speed up text-to-speech pipeline vocoding
• reconstruct audio from acoustic features in a single forward pass
• integrate a vocoder with Bark text-to-audio model
• train a custom vocoder on my own dataset

When to choose

• you need fast, high-quality waveform synthesis from mel-spectrograms or codec tokens
• you want a lightweight vocoder (under 14M parameters) for TTS pipelines
• CPU/GPU inference speed matters more than time-domain modeling

When to avoid

• you need sample rates or codecs beyond the provided 24kHz pretrained models
• you require time-domain waveform modeling or streaming low-latency synthesis
• you need a non-PyTorch framework

Facets

library · maturity stable

audio-processing machine-learning deep-learning llm-inference speech-processing machine-learning deep-learning python cross-platform vocoder gan text-to-speech audio-synthesis mel-spectrogram encodec pytorch tts audio gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="gemelo-ai/vocos")