README.md

---
license: mit
tags:
- audio
- speech
- autoencoder
---
# Versa AE v2 — Audio Samples

v2 autoencoder with a compact MPD (41M→2.7M), lower discriminator LR (5×10⁻⁵), and higher feature matching weight (λ_fm=5.0). Trained from step 100K→150K on LJSpeech + LibriTTS (~944h).

<table><tr><th>Original</th><th style="background-color:#e6f7ef">🔥 Versa AE v2</th><th>Vocos</th></tr>
<tr><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/5895_34615_000016_000000_orig_44k.wav" type="audio/wav"></audio></td><td style="background-color:#e6f7ef"><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/5895_34615_000016_000000_versa.wav" type="audio/wav"></audio></td><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/5895_34615_000016_000000_vocos.wav" type="audio/wav"></audio></td></tr>
<tr><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1993_147964_000003_000001_orig_44k.wav" type="audio/wav"></audio></td><td style="background-color:#e6f7ef"><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1993_147964_000003_000001_versa.wav" type="audio/wav"></audio></td><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1993_147964_000003_000001_vocos.wav" type="audio/wav"></audio></td></tr>
<tr><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1272_141231_000013_000001_orig_44k.wav" type="audio/wav"></audio></td><td style="background-color:#e6f7ef"><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1272_141231_000013_000001_versa.wav" type="audio/wav"></audio></td><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1272_141231_000013_000001_vocos.wav" type="audio/wav"></audio></td></tr>
<tr><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1988_147956_000014_000009_orig_44k.wav" type="audio/wav"></audio></td><td style="background-color:#e6f7ef"><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1988_147956_000014_000009_versa.wav" type="audio/wav"></audio></td><td><audio controls preload="none" style="width:150px"><source src="https://huggingface.co/datasets/danivs10/versa-ae-samples/resolve/main/samples/1988_147956_000014_000009_vocos.wav" type="audio/wav"></audio></td></tr></table>

Samples from LibriTTS dev-clean, resampled to 44.1kHz with torchaudio. Versa reconstructions use the step 150K checkpoint with EMA weights (Mel L1≈0.50). Vocos reference uses the pretrained Vocos checkpoint.

| Metric | v1 (step 100K, 41M MPD) | v2 (step 150K, 2.7M MPD) |
|---|---|---|
| MPD size | 41M | **2.7M** |
| D loss | 2.79 | **3.2** (balanced) |