Autoencoder
Hello!
The artifact description in the card (broadband hiss, hum bleeding in) sounded familiar. I already ran a quick test: pure encode -> decode round-trips through the VAE alone, no DiT, no LoRA, on close-mic'd broadband noise/texture content (ear massage, tapping, scratching etc.). Both stable-audio-open-1.0's own autoencoder and Stable Audio 3's SAME autoencoder lose that kind of texture almost entirely on reconstruction. It gets replaced with quieter, smoother, reverb-y output - with zero training involved. Have you tested the autoencoder in isolation, or was the "dataset labeling/trigger-class separation" diagnosis based on the full pipeline? Wondering if the autoencoder itself might be a bigger factor than the labels for this content type.
Thanks, this is a really good point and we hadn't tested it in isolation before. The "labeling" line in the card was a hypothesis, not a measured result. I just ran the same check: encode → decode through stable-audio-open-1.0's VAE alone (no DiT, no LoRA) on 26 of our trigger clips, 12 of them close-mic'd YouTube segments from the v1 training set (tapping, scratching, crinkling, page turning, etc.). On our data the texture largely survives. Energy per band stays within ~±2 dB, including 4–20 kHz. High-frequency transient flux keeps 92–105% of the original. Log-spectral distance is about the same as a white-noise control, so it's mostly the floor for resynthesising noise-like content. There's some smoothing of the finest HF detail, but not the quieter, reverb-y replacement you describe. The v1 artifacts were also other sounds bleeding in (hum, snoring-like FX, mouth clicks), which fits weak conditioning labels better than a lossy codec. I haven't tried SA3's SAME autoencoder yet. Would you share a clip and your round-trip code? I'm wondering if sample rate, mono/stereo, input scaling or chunked decoding explains the difference. If the VAE does turn out to be the bottleneck for some content, fine-tuning the decoder on ASMR is the next thing I'd try.