Product and requirements

Product boundary

insonic owns orchestration, user configuration, the library catalog, processing history and the presentation of source-linked results. It integrates media tools, Cueson, model runtimes, storage and databases through explicit adapters. Local processing and hosted providers are both supported routes. Speaker-tagged audio segments across the library can form a user-elected training corpus; trained artifacts belong to their originating speaker records. Training is optional and never a prerequisite to using the library.

Users can import local files or acquire user-targeted remote sources through acquisition adapters. Acquisition adapters declare supported protocols and services; failures explain unsupported formats or authentication requirements.

The primary journey is Add media, Choose processing, Process, Explore. A user can begin without learning model servers, database syntax or pipeline internals. Advanced users can select models, endpoints, channels, storage/database backends, adapters and query parameters. The CLI defines core application operations, and the GUI wraps those operations with a parity goal; explicitly documented advanced or experimental capabilities may reach the GUI later.

Requirement map

Stable requirement IDs identify observable product behavior. The named contracts define acceptance.

Completion and exceptions

Speech-capable input does not require supplied subtitles. Successful processing ends with a valid correlated transcript and an exportable subtitle file. Pending, cancelled and failed processing remains visible and resumable; it is never labelled complete. Music and ambient sound classifications may skip automatic speech work. Mixed media can still contain speech, and a user can force transcription or override a mistaken classification. Neither a model's classification nor the absence of speech confidence makes a file ineligible for the library.

Keep uncertain speaker identities and uncertain calendar dates usable. The application records uncertainty explicitly instead of guessing. Direct querying and exploration work without AI query assistance enabled.

Configured routines run without repeated approvals. Only the affected operation requires additional input when a date cannot be interpreted, a destructive effect is not already authorized, or a new external processing route has not been selected. Confidence filters and attribution policies are user-configurable automation inputs; uncertain results remain labelled and correctable without mandatory review queues.

Delivery dependencies

Runtime, storage and media contracts establish the import-to-subtitle workflow on all three operating systems, with full S3-compatible storage and PostgreSQL support. Provider flexibility and speaker continuity support graph extraction, desktop exploration and optional speaker training. Model discovery and retrieval begin with CLI contracts; desktop presentation follows the shared operation model. See delivery outcomes for dependencies. Multi-user collaboration and an insonic-hosted library service are outside these application contracts.