Frontier · Sub-page
Current vs aspirational.
Three buckets. Examples per bucket. If a claim about crows isn't in here, treat it with skepticism until it is.
AI narration · Frontier · Current vs aspirational
Five versions, named honestly. v0 is the live site you're reading now — interactive atlas, real audio for seven of nine clusters, narrator everywhere, no live similarity search yet. v1 adds Perch 2.0 embeddings on a real production corpus and a similarity-search panel. v2 adds upload-your-own-crow with cluster placement. v3 extends to other corvid species. v4 brings the dataset to Hugging Face. v5 is real-time bidirectional synthesis — the speculative end-state most popular coverage assumes already exists. We don't promise it. The frontier page exists so you can see exactly what's shipping versus what's still aspiration.
Demonstrated
- Automatic detection of crow vocalizations in long recordings (BirdNET, Perch, NatureLM-audio).
- Unsupervised discovery of call categories from raw audio.
- Caller sex and individual identity inference from a single call.
- Behavioral-context probabilities per cluster, with synchronized observation.
- Zero-shot species and behavior classification via NatureLM-audio prompts.
- Measurable group-level acoustic centroid differences (dialect signal).
Emerging
- Cross-population generalization without per-region fine-tuning.
- Compositional / sequence-level decoding (statistical evidence, weak behavioral evidence).
- Real-time on-device embedding for field use.
- Continuous repertoire mapping across multi-year datasets.
- Functional dialect tests via cross-group playback at scale.
Not yet science
- A 'crow dictionary' with human-language glosses.
- Generating novel crow utterances with predictable semantic effects.
- Real-time bidirectional dialogue between human and crow.
- Reliable individual-translation across never-recorded crows.
- Any claim that we 'understand' what crows are saying.
The line between buckets moves. Some "emerging" capabilities (graded-call clustering, individual identity from SSL embeddings) crossed into "demonstrated" over the last 24 months. Some "not yet science" items (compositional decoding) might cross into "emerging" in the next 24, if wearable-logger datasets keep growing and ethically defensible playback protocols mature.
CrowLingo updates this page when a credible peer-reviewed demonstration shifts the line. We don't update it for press releases.