Project goals

Concrete, executable goals that deliver on the mission.

G.1: Establish a universal interface

G.1.1: Standardize the core interface. Define a near-zero-dependency core protocol (StandardASR) that normalizes transcribe / transcribe_async (batch) and start_transcription (streaming).
G.1.2: Audio input negotiation and a constant result. A unified audio-input type system (AudioInput discriminated union) with deterministic negotiation. Lossy steps surface as structured diagnostics; impossible conversions fail loudly. The result is a constant-schema TranscriptionResult whose shape never changes with parameters.
G.1.3: Properties, capabilities, metadata, and config declarations. Engines make four machine-readable declarations: Properties (static I/O identity), Capabilities (hierarchical feature tree), Metadata (typed lifecycle and operational facts), and Config (typed, UI-renderable Pydantic model with secret-field marking).
G.1.4: Standardize optional features. Streaming, word timestamps, diarization, phrase hints, and other advanced features have standard interfaces, standard return formats, and a fail-closed capability query (supports()).
G.1.5: Unify streaming semantics. A single event protocol (partial / final / supersede / progress / done / error) with segment lifecycle and explicit stability guarantees covers every real-world streaming behavior -- from rewriting interims to append-only token streams to engines that re-cut segments they already sent.

G.2.1: One-command compliance. A compliance test suite that engine authors run with standard-asr compliance run to verify their implementation.
G.2.2: Ready-to-use tooling. A CLI for discovery and quick transcription, a FastAPI server that exposes any engine over HTTP/WebSocket, and a dependency conflict doctor.
G.2.3: Boilerplate templates. A reference plugin template for quick starts.

G.3.1: Dynamic config generation. All engine parameters are Pydantic models with JSON Schema output and secret-field marking, so UIs can render config forms without instantiating the engine.
G.3.2: Plugin auto-discovery. Entry-point-based discovery: install a plugin, and it appears in discover_models() with no application-side configuration.

G.4.1: Core and implementation separated. The core package is a near-zero-dependency protocol (numpy + pydantic). Each engine lives in its own independently maintained, independently licensed package.
G.4.2: Dependency and license isolation. Plugin architecture keeps conflicting dependencies and restrictive licenses contained. For hard conflicts (for example, numpy 1.x vs 2.x), standard-asr doctor diagnoses and process isolation is the escape hatch.
G.4.3: Plugin catalog. A public directory of compliant engines (known plugins, capability summaries, licenses).

G.5.1: Wire contract as first-class spec. The HTTP/WebSocket contract evolves into an independently versioned, language-neutral specification. The Python FastAPI server is a reference implementation, not the definition.
G.5.2: Two-layer isomorphism. The in-process Python protocol and the wire protocol share the same capability and declared-metadata models, result schema, and event semantics. Any evolution in one layer must have a corresponding expression in the other. Per-model metadata uses a dedicated endpoint so bulk discovery remains import-free.