Design a service that identifies a recorded song from a short, noisy audio sample. The system stores a catalog of reference tracks. A client uploads or streams a brief clip, and the service returns the best matching track with a confidence score or reports no reliable match. Discuss reference preprocessing, fingerprint indexing, query matching, and horizontal scaling. ### Constraints & Assumptions - Clips may start at any offset and may include background noise or compression artifacts. - The same recording should match despite moderate volume changes. - False positive and no-match behavior must be measurable and tunable. ### Clarifying Questions to Ask - How long are typical clips and how quickly is a result required? - Is the catalog batch-loaded or updated continuously? - Do we identify the exact recording or merely the underlying composition? ```hint Preserve relative time Individual spectral landmarks are common; pairs of landmarks plus their time difference are much more selective. ``` ### What a Strong Answer Covers - Robust audio fingerprint extraction and an inverted index over fingerprint keys. - Candidate generation followed by alignment voting on a consistent time offset. - Confidence thresholds, no-match results, catalog updates, partitioning, and observability. ### Follow-up Questions - How would you remove a track without rebuilding the full index? - What causes false positives, and how would you select a threshold? - How would live microphone streaming change the API?