V2A Doppler scores — car passby set

Does the pitch of the generated sound follow the car the way a real passby does (higher while approaching, lower while receding)? A rendered car drives past a fixed camera; LTX-2 and Cosmos3 generate the sound, 10 seeds per input, all kept. Every number below is one generated clip — click it to watch that exact video with its spectrogram and the pitch track the estimator measured. See also: distance scores · azimuth scores.

Estimator (one fixed setting for every clip, real or generated). Mono mix, 44.1 kHz. Log-frequency spectrogram of 60–400 Hz (240 bins/octave, 186 ms window, 20 ms hop), keeping only fine spectral structure. Every frame within ±1.5 s of a window centre is cross-correlated against a de-warped template, giving a pitch-shift track φ(t) in cents. A retarded-time passby model (speed v, closest distance d, closest-approach time t0, offset) is fitted to φ(t) with a robust loss. The reading is v. Fit bounds: v 1.08–216 km/h, d 0.5–40 m, t0 within ±0.6 s of the centre. Code: flowlib.estimate_lowband.

Primary setting = low-band, auto centre (centre = loudness peak, no ground truth). This is the only setting the Oct 14 run was scored with. The Oct 9 run also has low-band, centre at the video crossing time (2.5 s) and a supplementary 300–1800 Hz template. The true speed and distance never enter any fit.

How to read it. Error = reading − true speed. Pinned = the fit ended on a speed or distance bound, so the number is an artefact of the bound, not a measurement. There is no acceptance gate: a low fit RMS does not make a reading trustworthy (the Oct 9 arc control, which has no Doppler at all, gets low-RMS fits at 30–90 km/h). Grey numbers = what the same statistic reaches by chance, after shuffling which true speed goes with which input (20 000 shuffles).

1 · Does the estimator work on real recordings?

The same estimator, same setting, on real iPhone passbys (speed reported by the driver, approximate) and the original doppler_gt horn clip (reference from the horn's own frequency ratio, not GPS). Centre: auto as above, or video = the hand-marked time the car body crosses the image centre (±0.1 s). Readings were made before the speeds were known. The recordings themselves are not published here; click a reading to see its spectrogram and pitch track.

2 · Generated audio

Same text for every input and both models, with no speed or Doppler hint:  negative:

Summary

one clip (one seed) pinned at a fit bound true speed

3 · By closest distance and direction

Each cell is one input (10 seeds): median reading, interquartile range, median fitted closest distance and how many of the 10 fits are pinned. The estimator fits speed and distance jointly, so a correct reading should not depend on distance or direction; the true distance is in the row label.

4 · Picking a seed

5 · Per-clip readings

close to true speed off by ≥ 50% pinned at a fit bound arc row: false Doppler (darker = larger)

km/h. Shift/Ctrl/Cmd-click a number to open the raw .mp4 in a new tab. Files: dataset folder (generated videos, per-clip figures, scores.json, predictions.csv), tables per_clip_readings.csv, summary.csv. Pitch tracks and spectrograms on this page are drawn from the saved low-band arrays; the drawn fit curve uses the saved v, d, t0.