Knowledge Base

Problem

Users can clearly see ultrasonic vocalizations (USVs) in the UltraVox spectrogram view, but the automated detection function fails to reliably identify or classify the calls. Detection either misses calls entirely, over-detects noise as calls, or only partially captures vocalizations.

Details

Affects UltraVox XT 3.2 and UltraVox XT 4.x. Particularly reported for neonatal mouse pup calls (P7 age), which are short and variable in structure. Auto-generated detection parameters from a single selected call are frequently too narrow or imprecise to generalize across a recording. High gain settings, background noise, and improper parameter tuning are common contributing factors. The problem is parameter configuration, not a software defect.

Solution

1. Start with the cleanest recording available — minimize background noise before adjusting detection parameters.
2. Apply a high-pass filter in UltraVox to remove low-frequency noise that may interfere with call boundaries.
3. Set the software Gain as low as practical (approximately 20) so that vocalizations are clear without amplifying the noise floor. High gain can cause clipping or noise artifacts that disrupt detection.
4. In the spectrogram view, draw a selection box tightly around one representative, complete call — avoid including silence or noise at the edges.
5. Use the auto-generated parameters from this selection as a starting point only; do not treat them as final values.
6. Adjust the minimum and maximum frequency bounds to bracket the typical frequency range of the calls in your recording.
7. Tune the minimum duration, gap tolerance, and amplitude threshold parameters incrementally, re-running detection after each change and comparing results against the visible spectrogram.
8. Validate final parameters against multiple recordings before applying them to a full dataset.
9. Consult the UltraVox manual appendix for species- and age-specific parameter guidance as an additional reference point.