💡 Acoustic colony health screening via phone microphone during inspections

[github-actions[bot]]

Summary

During voice-first inspections, capture a 30-second entrance audio sample and analyze it using ML-based acoustic classification to surface colony-state indicators — queenright confidence, agitation level, and swarm-readiness — as evidence inputs to the recommendation engine. The phone mic is already active during inspections, making this near-zero marginal cost. This turns every smartphone into a non-invasive colony health sensor.

Market Signal

BeeVe paper (Hammami & Abdulaziz, May 2026) demonstrated unsupervised acoustic state discovery from unlabeled hive audio using PaSST + VQ-VAE, separating queenright vs queenless conditions with JSD 0.609-0.688 and identifying three distinct queenless sub-states — all without labeled training data. BuzzBox exists as dedicated hardware ($149+) but no app offers phone-mic-based acoustic screening. A comprehensive TinyML survey (Sucipto et al., Sept 2025) confirms on-device inference feasibility for bee health classification across four functional areas. The new UrBAN dataset provides standardized urban beehive acoustics for benchmarking.

User Signal

55.6% colony losses in 2024-2025 — the worst annual loss since tracking began in 2010. Newbie beekeepers struggle most with interpreting colony state during inspections. Acoustic signals provide an additional evidence layer that doesn't require visual interpretation expertise, sensor hardware investment, or even opening the hive. PRD already specifies FR56-FR58 for acoustic analysis but no tracking discussion exists.

Technical Opportunity

The architecture uses Gemini multimodal via Vertex AI (text+image+audio+video embeddings via Embedding 2.0). The voice-first inspection flow means the mic is already capturing audio — dual-purpose the audio stream for colony state analysis at near-zero marginal cost. Gemini 3.1 Pro handles audio classification natively with its 1M-token multimodal context. Google Chirp 3's built-in denoiser addresses field noise concerns. The async worker service can process audio post-capture without blocking the inspection flow.

Assessment

Dimension	Score	Rationale
Feasibility	high	BeeVe proves the ML approach works; Gemini handles audio natively; mic is already active during inspections
Impact	high	Sensor-grade colony intelligence without hardware; directly addresses newbie interpretation anxiety
Urgency	high	Differentiates from sensor-dependent competitors; worst colony loss year creates demand for accessible health monitoring

Adversarial Review

Strongest objection: Accuracy in real field conditions with wind, traffic, and other environmental noise will be far below controlled lab/research conditions, making the feature unreliable.

Rebuttal: The feature provides one additional signal in the confidence-scored recommendation engine, not a definitive diagnosis. Even a 70% queenright confidence indicator reduces uncertainty for newbies. Modern noise reduction (Google Chirp 3 denoiser) and controlled 30-second capture windows with user guidance ("hold phone near entrance") mitigate environmental noise. The mic is already active during voice inspections — this is incremental signal at near-zero marginal effort and cost.

Suggested Next Step

Prototype audio capture during a mock voice inspection session, send samples to Gemini Audio for classification, and evaluate accuracy against 10+ real hive recordings from the UrBAN public dataset. Define the acoustic evidence schema for integration into the recommendation engine's confidence scoring.