Can YAMNet Detect Unseen Sudden Sounds in Real Time? A 48-Stream Evaluation

Short summary

An experiment testing whether YAMNet embeddings can detect unseen sudden sounds in real time without pre-registering sound classes. The best configuration achieved 45.8% recall (22 of 48 events), processing faster than real time. Results did not support the hypothesis that simple distance over YAMNet embeddings reliably detects novel sounds, though the pipeline and code are openly available.

• •YAMNet streaming novelty detection achieved 45.8% recall on 48 events
• •Processing was faster than real time but detection reliability was insufficient
• •Full code and results available in open-source kiarina/labs repository

Generated with AI, which can make mistakes.