Assessment of spectrogram correlation, eigensounds, and YAMNet embeddings in detecting cattle vocalizations

• YAMNet embeddings achieved highest precision in cattle vocalization detection. • Two-stage pipeline: eigensounds (high recall) and YAMNet (high precision) • Eigensounds outperform spectrogram correlation for recall in dairy cow monitoring. • Spectrogram correlation indicative of individual-specific matching. • Web application prototype enables veterinarians to deploy models on recorded audio. Acoustic monitoring enables unattended detection of vocal events that may support health and welfare surveillance in Precision Livestock Farming (PLF). This study presents a systematic comparison of three approaches for automatic detection of cattle vocalizations, characterizing their trade-offs in terms of precision–recall balance, sensitivity to individual variability, and practical applicability. We used recordings from twenty Holstein multiparous lactating dairy cows, each monitored individually in a tied stanchion barn, with temporary isolation of 4 h post-morning milkings to facilitate data capture. We compared: (1) spectrogram cross-correlation using Praat and Parselmouth, (2) an eigensounds method based on Principal Component Analysis, and (3) embedding similarity using the pre-trained YAMNet model. On a curated set of twenty 1-hour recordings collected from isolated pens, the correlation method achieved rather modest results, with only three recordings achieving an F1-score above 0.5; performance was strongest when the template matched the target individual, with cross-individual templates yielding consistently lower scores. With suitable component counts and thresholds, the eigensounds approach yielded moderate results, with half of the recordings achieving an F1-score above 0.5, but high inter-individual variability, as adjusting thresholds improves precision for some subjects but greatly reduces recall for others. YAMNet achieved consistently high precision, with recall adjustable through thresholding and template selection, even under considerable noise. We further examine how thresholding, event merging, and duration capping influence precision–recall performance, and we outline a semi-automated labeling workflow, discussing the potential of wearable recorders and other strategies to enhance accuracy. These findings primarily inform method behavior under controlled single-source conditions and provide a reference point for future evaluation in more complex, multi-animal farm settings.