Segment duration
20, 40, 60, 80, 100, 200, 500 and 1000 ms.
Problem
Voice activity detection is usually evaluated on relatively clean, well-segmented audio. This benchmark instead measures how robust speech/non-speech decisions remain when acoustic evidence is deliberately degraded along controlled psychometric dimensions.
Additive noise
SNR levels from -10 dB to +20 dB.
Reverberation
RT60 values of 0.0, 0.3, 1.0 and 2.5 seconds.
Spectral filtering
None, bandpass, lowpass and highpass conditions.
Systems compared
| System | Balanced accuracy | Speech recall | Non-speech recall |
|---|---|---|---|
| Qwen2-Audio-7B + OPRO-LLM | 82.6% | 74.7% | 90.6% |
| Qwen2-Audio-7B + LoRA + OPRO-Template | 93.3% | 92.8% | 93.8% |
| Qwen3-Omni-30B frozen + hand prompt | 91.1% | 87.4% | 94.7% |
| Silero VAD | 88.9% | 78.8% | 99.1% |
The strongest balanced result came from the adapted 7B model rather than the larger frozen 30B model. Silero remained the most conservative system, with very high non-speech recall but lower speech recall.
What I built
Benchmark
Controlled degradation bank spanning duration, SNR, reverberation and spectral filtering.
Adaptation
LoRA fine-tuning for Qwen2-Audio-7B, including 4-bit quantization and PEFT-based training.
Prompt search
Hand prompts, OPRO-LLM and OPRO-Template treated as experimental factors. The prompt-optimization stage contains 435 evaluations over 71 unique prompts.
Evaluation
Balanced accuracy, class recall, per-condition breakdowns, bootstrap confidence intervals, McNemar tests and multi-seed analyses.
Artifacts
JSON metrics, CSV predictions, statistical reports, LaTeX tables, figures, audit reports and Slurm job support.
Robustness findings
Short evidence
LoRA + OPRO-Template reaches DT90 = 96 ms, approximately the amount of audio required to reach 90% balanced accuracy.
Noise
LoRA adaptation produces a more stable robustness profile under additive noise than prompt optimization alone.
Failure modes
Human and animal vocalizations are important confounders. Laughing, coughing and crying-baby clips remain difficult non-speech cases.
Model size
The experiment provides a concrete counterexample to the assumption that the largest frozen audio-language model necessarily gives the strongest VAD result.
Evidence and reproducibility
The public repository contains the experiment configuration, evaluation scripts, fine-tuning code, prompt-optimization scripts, consolidated metrics, statistical analyses, figures, LaTeX tables and audit reports used to reconstruct the headline results.