First, does the sound resemble a bird?
BirdTalk reads short audio windows on device and measures loudness, peak level, zero-crossing rate, spectral centroid, high-frequency energy, and bird-likeness. Live listening and imported audio follow the same analysis path.
Then form a limited state hypothesis
Current results include calm, contact, attention seeking, activity, unease, alarm, human-speech mimicry, and uncertain. The interpreter combines features within these limited categories instead of inventing a human sentence.
Evidence strength is not accuracy
The percentage describes how clearly the audio, bird-like characteristics, and repetition support the current rule. It has not been probability-calibrated on a large independent labeled dataset, so 70% does not mean the conclusion is scientifically 70% likely to be correct.
When evidence is weak, clues conflict, or the room is too noisy, the honest result is “uncertain.”
Your observation remains essential
The same call can mean something different when a bird is alone, greeting you, or startled. Check posture, environment, eating, breathing, and what happens next; use the journal to record context.
Automatic comfort has hard limits
Playback is possible only when you explicitly enable it, a recording exists for the current scene, the broad state and evidence meet the safety policy, and quiet hours, cooldown, and daily limits allow it. Analysis pauses during playback to prevent a speaker–microphone loop.
When to contact a veterinarian
BirdTalk cannot identify illness. If your bird persistently open-mouth breathes, shows pronounced tail bobbing while breathing, crashes around the cage, becomes unusually lethargic, stops eating, or shows another physical abnormality, stop audio interaction and contact an avian veterinarian.