Sarama studies non-human communication at consumer scale. We spent years decoding whales and dolphins. We still could not understand our own dogs. The collar is the unlock. The lab is the point.
Animal communication has been bottlenecked by data, not by AI. No bark means the same thing across all dogs in all contexts. Meaning is contextual, causal, and individual. The path forward is multimodal, longitudinal, naturalistic data from tens of thousands of animals living their normal lives.
YAMNet-based segmentation, MFCC embeddings, and unsupervised clustering to recover per-dog vocal repertoires, not a universal bark-to-English lookup.
Time-aligned signals on a single device: directional microphone, on-device visual features, IMU, and physiological proxies. Meaning lives in the joint distribution.
Acoustic fingerprinting, individual-ID models, and routine-aware context. Reliable interpretation of one dog is built from that dog's history, not a generic prior.
Inference compressed to ARM Cortex-M class hardware. Devices transmit features, not raw audio or video, so household-scale data collection is ethically tractable.
The earlier app taught us labels reflect dog parent expectation, not dog meaning. Ground truth now comes from sensor-grounded signals; dog parent feedback is treated as soft, biased prior.
The pivot from bark-labeling to sensor-grounded modeling is on the record. We publish methods, datasets, and what didn't work, because the science isn't settled.
Tens of thousands of collars entering ordinary homes is the first place where interspecies communication becomes empirical. Dogs first. Methodology general.