🤗Hugging Face | KVAE GitHub | Habr Article | Project Page | Technical Report (soon)

KVAE-Audio

KVAE-Audio is a continuous, full-band (48 kHz) audio autoencoder. It compresses raw waveforms into compact continuous latents and reconstructs them with high fidelity across speech, music, and general sound. The model is designed not only for faithful reconstruction, but as a latent space for generative models — in our internal text-to-audio pipeline, swapping the autoencoder for KVAE-Audio improves generation quality under a fixed generator.

Inference instruction

Setup

First install PyTorch (adjust for your CUDA setup):

pip install torch==2.11.0 torchvision==0.26.0 torchaudio==2.11.0 --index-url https://download.pytorch.org/whl/cu128

Then clone the repository and install KVAE-Audio from the local package:

git clone https://github.com/kandinskylab/kvae-audio.git
pip install -e ./kvae-audio

Tested with Python 3.14.

Download weights

Checkpoint kvae-audio.pt is hosted on Hugging Face: kandinskylab/KVAE-Audio.

pip install huggingface_hub
hf download kandinskylab/KVAE-Audio kvae-audio.pt --local-dir .

Or from Python:

from huggingface_hub import hf_hub_download

weights_path = hf_hub_download("kandinskylab/KVAE-Audio", "kvae-audio.pt")

Run inference

Place input wavs in input_samples. Reconstructions are written to output_samples.

python scripts/infer.py \
  --weights_path kvae-audio.pt \
  --input_path input_samples \
  --output_path output_samples

Evaluate reconstructions

Compare reference wavs in input_samples against reconstructions in output_samples (same filenames). Writes metrics.csv to output_samples.

python scripts/eval.py \
  --input_path input_samples \
  --output_path output_samples

Python API

By default encode returns the mean latent (mu); pass sample=True for stochastic sampling. For end-to-end reconstruction, model(waveform, sample_rate)["audio"] is equivalent.

import torch
import soundfile as sf
from kvae_audio import KVAEAudio

model = KVAEAudio.load("kvae-audio.pt", map_location="cpu")
model.eval()

data, sr = sf.read("audio.wav")
waveform = torch.from_numpy(data).float().unsqueeze(0).unsqueeze(0)
length = waveform.shape[-1]

with torch.no_grad():
    latents, _, _, _ = model.encode(waveform, sample_rate=sr)
    audio = model.decode(latents)[..., :length]

sf.write("output.wav", audio.squeeze(0).T.numpy(), sr)

Evaluation results

Generative quality is established under a fixed generator — same DiT architecture, training data, and number of steps — varying only the autoencoder. We report objective generation metrics and blind human side-by-side below.

Evaluation of latent space qualities for generation

Audio Generation Examples

Below are qualitative examples generated from the same text prompts using four different models.

Example 1 (speech)

Prompt:

Low and gravelly, with a southern Russian accent, a man says slowly, <S>Я говорил тебе, что так и будет<E>. A sharp inhale precedes the line. The space is acoustically dry, with minimal room tone and a slight hiss.

KVAE-Audio Details kvae1.mp4 Download

MMAudio Details mma1.mp4 Download

DACVAE MovieGen Details moviegen1.mp4 Download

SAME-L Details samel1.mp4 Download

---

Example 2 (background)

Prompt:

In a home kitchen, oil sizzles in a pan, a knife chops on a board, and a woman hums softly in Russian, a refrigerator humming beneath it all. Small kitchen, light reverberation. High-fidelity.

KVAE-Audio Details kvae2.mp4 Download

MMAudio Details mma2.mp4 Download

DACVAE MovieGen Details moviegen2.mp4 Download

SAME-L Details samel2.mp4 Download

---

Example 3 (music)

Prompt:

In a large reverberant hall, a brass band launches into a lively march. Trumpets and cornets carry the melody while trombones drive the harmony, brisk and precise. Clean recording, natural reverb, faint equipment hiss.

KVAE-Audio Details kvae3.mp4 Download

MMAudio Details mma3.mp4 Download

DACVAE MovieGen Details moviegen3.mp4 Download

SAME-L Details samel3.mp4 Download

---

AudioCaps test set

Model	# Params	Latent dim	CLAP↑	CE↑	PQ↑	FAD (PANNs)↓	FAD (PASST)↓	FAD (VGGIsh)↓
MMAudio 44.1kHz	427.6M	40	0,336	3,909	6,192	17,873	195,910	1,364
DACVAE MovieGen	107.7M	128	0,313	3,772	6,167	20,558	234,312	1,700
SAME-L	852.1M	256	0,322	3,588	5,756	18,446	240,635	1,325
KVAE-Audio	166.9M	64	0,344	3,982	6,242	15,381	193,760	1,210

Song Describer

Model	# Params	Latent dim	CLAP↑	CE↑	PQ↑	FAD (PANNs)↓	FAD (PASST)↓	FAD (VGGIsh)↓
MMAudio 44.1kHz	427.6M	40	0,356	7,136	7,707	5,412	158,599	0,356
DACVAE MovieGen	107.7M	128	0,312	6,953	7,538	10,194	214,009	1,046
SAME-L	852.1M	256	0,345	7,076	7,465	8,442	250,668	0,987
KVAE-Audio	166.9M	64	0,339	7,216	7,929	7,971	189,427	0,599

LibriSpeech test-clean

Model	# Params	Latent dim	CLAP↑	CE↑	PQ↑	FAD (PANNs)↓	FAD (PASST)↓	FAD (VGGIsh)↓	WER↓	CER↓
MMAudio 44.1kHz	427.6M	40	0,368	5,704	6,629	8,305	105,931	2,001	0,257	0,593
DACVAE MovieGen	107.7M	128	0,413	5,482	7,052	5,008	210,478	1,501	0,911	1,048
SAME-L	852.1M	256	0,379	4,617	5,024	10,257	301,508	2,721	0,349	0,629
KVAE-Audio	166.9M	64	0,389	5,906	6,940	4,677	185,609	2,138	0,244	0,576

Reconstructions

Reconstruction is evaluated on open datasets across domains (the released weights directly substantiate these numbers). Baselines: MMAudio 44.1 kHz VAE, DACVAE from MovieGen Audio, SAME-L (Stable Audio 3 VAE).

AudioSet eval

Model	# Params	Latent dim	MEL↓	STFT↓	Waveform↓	SI-SDR↑	SDR↑	SNR↑
MMAudio 44.1kHz	427.6M	40	0,636	1,938	0,106	-32,080	-2,682	-2,686
DACVAE MovieGen	107.7M	128	0,669	2,275	0,029	8,384	9,421	9,416
SAME-L	852.1M	256	0,986	2,726	0,027	9,586	10,347	10,339
KVAE-Audio	166.9M	64	0,537	1,770	0,027	9,065	9,920	9,933

MUSDB18-HQ

Model	# Params	Latent dim	MEL↓	STFT↓	Waveform↓	SI-SDR↑	SDR↑	SNR↑
MMAudio 44.1kHz	427.6M	40	0,681	1,865	0,114	-40,204	-3,274	-3,273
DACVAE MovieGen	107.7M	128	0,519	1,762	0,024	9,688	10,046	10,047
SAME-L	852.1M	256	0,668	1,786	0,023	10,278	10,648	10,648
KVAE-Audio	166.9M	64	0,516	1,725	0,022	10,390	10,675	10,677

EARS

Model	# Params	Latent dim	MEL↓	STFT↓	Waveform↓	SI-SDR↑	SDR↑	SNR↑	PESQ↑
MMAudio 44.1kHz	427.6M	40	0,616	1,395	0,030	-29,947	-2,728	-2,697	2,424
DACVAE MovieGen	107.7M	128	0,453	1,310	0,006	10,264	10,680	10,681	4,246
SAME-L	852.1M	256	0,774	1,575	0,007	9,939	10,374	10,376	2,982
KVAE-Audio	166.9M	64	0,463	1,314	0,006	9,952	10,377	10,384	4,266

Citation

@misc{kvae_audio_2026,
    author = {Ivan Kirillov, Denis Parkhomenko, Alexander Ivanov, Azat Saginbaev, Egor Silvestrov, Denis Dimitrov},
    title = {KVAE-Audio: a full-band continuous audio tokenizer for generative models},
    howpublished = {\url{https://github.com/kandinskylab/kvae-audio}},
    year = 2026
}

Package layout

kvae-audio/
├── kvae_audio/
│   ├── model/
│   │   ├── kvae_audio.py
│   │   └── base.py
│   ├── metrics/loss.py
│   └── nn/layers.py
└── scripts/
    ├── infer.py
    └── eval.py