Media intelligence / Featured system
Aural Echoes
A deterministic policy brain and broadcast architecture for a 24/7 AI-assisted station.
01 / Context
Context
A continuous station needs more than a playlist and a text generator. It has to choose music coherently, handle requests fairly, explain deferrals, respect rights, time voice around the recording, and keep broadcasting when a model or enrichment path is unavailable.
02 / Problem
Problem
Probabilistic generation sits inside a deterministic broadcast clock. The system must keep policy, library truth, timing, refunds, voting, and audio handoffs predictable while still using AI for analysis, voice, and adaptation.
03 / System
System
A TypeScript and SQLite station core reconciles read-only library sources, models tracks with musical and rights metadata, applies request and program policy, schedules analysis-aware Auto-DJ selections, generates constrained voiceovers, and hands production audio toward Liquidsoap and Icecast.
04 / Architecture
See the connective layer.
＋Architecture mode
System architecture / Aural Echoes
Library intelligence and audience state enter a deterministic policy brain; playout, voice, and rights decisions become validated audio handoffs to a production streaming layer.
- Read-only library and analysis sources
- Reconciled station metadata
- Program clock, request, and rights policy
- Auto-DJ selection and voice planning
- Validated playout handoff
- Liquidsoap / Icecast broadcast targets
05 / Hard parts
Where the work actually was
- 01
Keeping program-clock and request behavior deterministic while enrichment and TTS remain probabilistic.
- 02
Timing generated speech around detected vocal entry and validating duration before it reaches playout.
- 03
Reconciling media-library sources without mutating their source of truth.
- 04
Using lab evidence to replace an unsuitable custom audio-mixing path before calling the system production-ready.
06 / What changed or was learned
The field note
The most important outcome was architectural judgment. The policy brain and proof tooling held up; the custom production-audio path did not. Preserving that distinction made it possible to retain the valuable system while pivoting the broadcast layer.
07 / Evidence
Proof
Safe public evidence, translated from tests and runtime validation rather than raw internal logs.
Exercised the deterministic selection, request, program-clock, voiceover, and handoff paths in a lab runtime.
Selection decisions retain mood, energy, genre, era, tempo, key, confidence, request state, and rights context.
Voiceover planning accounts for detected vocal entry, fact-constrained prompts, generated duration, validation, and ducking envelopes.
Runtime proof exposed the limits of a custom audio mixer and informed a move toward Liquidsoap and Icecast for production audio.
08 / Stack
Technology in service of the system
- TypeScript
- SQLite
- Semantic HTML
- Custom CSS
- Browser Audio
- Hermes
- xAI
- TTS
- Liquidsoap
- Icecast
- FFmpeg
- Docker Compose
10 / Status and boundaries