This study reports the empirical reliability of Gemini models (2.5 Flash, 3.5 Flash, and 3.1 Pro) acting as audio judges that score full-duplex agent conversations directly from raw stereo waveforms.
- Gemini 2.5 Flash served as the ground-truth model, validated against three calibrated human raters on 209 stereo sessions across 8 production dimensions.
- On 5 of 8 dimensions, the LALM-human Spearman rho departed from pairwise human-human rho by at most 0.07, with overlapping confidence intervals on 7 of 8 dimensions.
- Gemini 3.5 Flash improved simple agreement to 8 of 8 dimensions, while 3.1 Pro rated several dimensions markedly lower than humans despite comparable rank correlation.
- The LALM is as sensitive as humans or better in 45 of 48 defect-dimension cells, though most are underpowered nulls rather than demonstrated parity.
- Human rating costs roughly two orders of magnitude more than the equivalent LALM workload for the evaluated cadence.
The data provides a defensible empirical basis for deploying the LALM as a substitute or fourth rater on dimensions where evidence supports it, though model swaps require specific calibration validation.