Audio VAE encode → decode roundtrip (highlighted targets, 2026-08-22 run)

Same path as encode_av_inmem.py / decode_av_inmem.py: load_audio_auto (44.1 kHz mono, 8 s) → ensure_max_of_audio → AudioAutoencoder(/group2/ct/weihanx/vae/vae) encode (latent 64×345) → decode. Two inputs per clip: the raw video audio (src) and the new highlighted wav from 8_sec_highlighted_0822 (target). Metrics are sample-aligned waveform SNR / Pearson corr / log-spectral distance (LSD) between VAE input and decoded output.

Summary (15 clips). Highlighted-target roundtrip: SNR median 13.8 dB (min 6.0, max 33.5), corr median 0.979; video-audio roundtrip SNR median 9.2 dB. RMS is preserved within 0.5 dB everywhere. Lowest-SNR cases are noisy/broadband clips (birds, engine, chickens, crowd, drum kit) — waveform SNR penalises phase differences, so listen to judge; the highlighted target generally round-trips slightly better than the raw mix because the foreground dominates.
idpromptsrc SNRsrc corrsrc LSDtarget SNRtarget corrtarget LSDtarget RMS in→out
Bm3Lfx58wvo_000200train3.870.798.2214.270.9818.05-14.6 → -14.5
Bm6puindzscbirds12.320.9713.256.560.88810.57-40.6 → -40.7
Bm8mxpOSmjw_000070melodic percussion instrument25.530.9998.3930.421.09.59-11.8 → -11.9
BmBido0MLFkidling engine5.060.84111.76.080.8769.89-16.9 → -17.0
BmFzGKRojvAbrass instruments32.561.08.7233.491.08.8-14.5 → -14.5
PifhBU-mSVA_000010infant15.330.9869.031.351.08.97-12.5 → -12.6
TOBw7DgXbgg_000047musician14.330.9878.5723.270.9987.73-12.7 → -12.8
xc74uXzCRWk_000013harmonica9.00.93911.8510.910.969.87-17.6 → -17.5
U-L9YCIdLbg_000080electric guitar14.390.98210.7716.130.9887.88-16.8 → -16.7
R93J7u5EWvkdrum kit1.660.65710.567.960.92110.36-18.0 → -17.9
Kz_gp7UE9IY_000001dog20.610.9966.6721.580.99710.3-12.6 → -12.9
mT49XQjXEvY_000000chickens6.450.8828.886.00.8710.1-16.7 → -17.1
hqvqPITWcrA_000260siren9.210.9417.938.630.9329.7-17.9 → -17.9
CsMMU1p3Sik_000030crowd5.510.85910.657.00.8998.1-13.0 → -13.2
iTdlEJPKT94_000300train horn6.840.89512.6213.850.97910.44-22.7 → -22.9

Bm3Lfx58wvo_000200 — “train” (first5)

video audio roundtrip: SNR 3.87 dB · corr 0.79 · LSD 8.22 dB  |  highlighted roundtrip: SNR 14.27 dB · corr 0.981 · LSD 8.05 dB · RMS -14.6→-14.5 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

Bm6puindzsc — “birds” (first5)

video audio roundtrip: SNR 12.32 dB · corr 0.97 · LSD 13.25 dB  |  highlighted roundtrip: SNR 6.56 dB · corr 0.888 · LSD 10.57 dB · RMS -40.6→-40.7 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

Bm8mxpOSmjw_000070 — “melodic percussion instrument” (first5)

video audio roundtrip: SNR 25.53 dB · corr 0.999 · LSD 8.39 dB  |  highlighted roundtrip: SNR 30.42 dB · corr 1.0 · LSD 9.59 dB · RMS -11.8→-11.9 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

BmBido0MLFk — “idling engine” (first5)

video audio roundtrip: SNR 5.06 dB · corr 0.841 · LSD 11.7 dB  |  highlighted roundtrip: SNR 6.08 dB · corr 0.876 · LSD 9.89 dB · RMS -16.9→-17.0 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

BmFzGKRojvA — “brass instruments” (first5)

video audio roundtrip: SNR 32.56 dB · corr 1.0 · LSD 8.72 dB  |  highlighted roundtrip: SNR 33.49 dB · corr 1.0 · LSD 8.8 dB · RMS -14.5→-14.5 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

PifhBU-mSVA_000010 — “infant” (multi-source)

video audio roundtrip: SNR 15.33 dB · corr 0.986 · LSD 9.0 dB  |  highlighted roundtrip: SNR 31.35 dB · corr 1.0 · LSD 8.97 dB · RMS -12.5→-12.6 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

TOBw7DgXbgg_000047 — “musician” (multi-source)

video audio roundtrip: SNR 14.33 dB · corr 0.987 · LSD 8.57 dB  |  highlighted roundtrip: SNR 23.27 dB · corr 0.998 · LSD 7.73 dB · RMS -12.7→-12.8 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

xc74uXzCRWk_000013 — “harmonica” (multi-source)

video audio roundtrip: SNR 9.0 dB · corr 0.939 · LSD 11.85 dB  |  highlighted roundtrip: SNR 10.91 dB · corr 0.96 · LSD 9.87 dB · RMS -17.6→-17.5 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

U-L9YCIdLbg_000080 — “electric guitar” (multi-source)

video audio roundtrip: SNR 14.39 dB · corr 0.982 · LSD 10.77 dB  |  highlighted roundtrip: SNR 16.13 dB · corr 0.988 · LSD 7.88 dB · RMS -16.8→-16.7 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

R93J7u5EWvk — “drum kit” (multi-source)

video audio roundtrip: SNR 1.66 dB · corr 0.657 · LSD 10.56 dB  |  highlighted roundtrip: SNR 7.96 dB · corr 0.921 · LSD 10.36 dB · RMS -18.0→-17.9 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

Kz_gp7UE9IY_000001 — “dog” (multi-source)

video audio roundtrip: SNR 20.61 dB · corr 0.996 · LSD 6.67 dB  |  highlighted roundtrip: SNR 21.58 dB · corr 0.997 · LSD 10.3 dB · RMS -12.6→-12.9 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

mT49XQjXEvY_000000 — “chickens” (multi-source)

video audio roundtrip: SNR 6.45 dB · corr 0.882 · LSD 8.88 dB  |  highlighted roundtrip: SNR 6.0 dB · corr 0.87 · LSD 10.1 dB · RMS -16.7→-17.1 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

hqvqPITWcrA_000260 — “siren” (multi-source)

video audio roundtrip: SNR 9.21 dB · corr 0.941 · LSD 7.93 dB  |  highlighted roundtrip: SNR 8.63 dB · corr 0.932 · LSD 9.7 dB · RMS -17.9→-17.9 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

CsMMU1p3Sik_000030 — “crowd” (multi-source)

video audio roundtrip: SNR 5.51 dB · corr 0.859 · LSD 10.65 dB  |  highlighted roundtrip: SNR 7.0 dB · corr 0.899 · LSD 8.1 dB · RMS -13.0→-13.2 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

iTdlEJPKT94_000300 — “train horn” (multi-source)

video audio roundtrip: SNR 6.84 dB · corr 0.895 · LSD 12.62 dB  |  highlighted roundtrip: SNR 13.85 dB · corr 0.979 · LSD 10.44 dB · RMS -22.7→-22.9 dBFS

Original video audio (VAE input)
Video audio → encode → decode
Highlighted target (VAE input)
Highlighted → encode → decode

Built by /group2/ct/weihanx/audio_hl_demo/audio_roundtrip.py + build_roundtrip_space.py