Kandinsky Lab Open-Sources KVAE-Audio: A High-Fidelity 48kHz Audio Autoencoder
Kandinsky Lab (Sber) releases KVAE-Audio, a 166.9M parameter continuous audio autoencoder achieving state-of-the-art reconstruction quality across speech, music, and general sound at 48kHz full bandwidth.
Overview
KVAE-Audio is a 166.9M parameter continuous audio autoencoder operating at 48 kHz full bandwidth. It compresses raw audio waveforms into a 64-dimensional continuous latent space and reconstructs them with high fidelity across speech, music, and general sound domains. The model is designed as a building block for generative audio pipelines — swapping the autoencoder alone improves generation quality under a fixed generator.
Key Results
KVAE-Audio achieves state-of-the-art or competitive results across multiple domains and metrics:
- 166.9M parameters — significantly smaller than MMAudio VAE (427.6M), DACVAE MovieGen (107.7M), and SAME-L (852.1M)
- 64-dim latent space — compact representation ideal for diffusion model integration
- Outperforms baselines on AudioSet, MUSDB18-HQ, EARS, AudioCaps, Song Describer, and LibriSpeech benchmarks
- Best overall reconstruction: lowest MEL, STFT, and Waveform distances on AudioSet and MUSDB18-HQ
- Best generation quality under fixed generator: highest CLAP score (0.344) and FAD scores on AudioCaps
Reconstruction Quality (AudioSet)
| Model | Params | Latent Dim | MEL↓ | SI-SDR↑ | SDR↑ |
|---|---|---|---|---|---|
| MMAudio 44.1kHz VAE | 427.6M | 40 | 0.636 | -32.08 | -2.68 |
| DACVAE MovieGen | 107.7M | 128 | 0.669 | 8.38 | 9.42 |
| SAME-L (Stable Audio 3 VAE) | 852.1M | 256 | 0.986 | 9.59 | 10.35 |
| KVAE-Audio | 166.9M | 64 | 0.537 | 9.07 | 9.92 |
Generation Quality (AudioCaps)
Under a fixed DiT generator, KVAE-Audio achieves the best CLAP score (0.344), cross-entropy (3.982), and perceptual quality (6.242), outperforming all baselines including the much larger SAME-L.
Architecture
KVAE-Audio is a continuous VAE (not discrete like DAC/EnCodec) designed to produce smooth, information-rich latents suitable for diffusion model training. Key architectural details include:
- Full-band 48kHz operation — no band-splitting or subband coding
- Continuous latent space — 64-dim, designed for flow-matching and diffusion backbones
- Lightweight decoder — efficient enough for real-time applications
Availability
The model is released under the MIT license on HuggingFace with pre-trained weights. The repository includes:
- Pre-trained VAE encoder and decoder weights
- Inference scripts
- Evaluation benchmarks
GitHub: https://github.com/kandinskylab/kvae-audio HuggingFace: https://huggingface.co/kandinskylab/KVAE-Audio
Significance
KVAE-Audio represents an important step for the open-source audio generation ecosystem. Its compact latent space (64-dim vs. 128-256 for competitors) combined with superior reconstruction quality makes it an attractive drop-in replacement for existing audio VAEs in text-to-audio, video-to-audio, and music generation pipelines. The MIT license further lowers the barrier for integration into both research and commercial projects.
Comments
Sign in with GitHub to join the discussion.