Audio production

From approved script to mastered episode

The intended route from approved segments through speech, mixing, loudness, and export. Almost none of it is built, and this page says so before it describes any of it.

The synthesis stages

Seven steps between an approved script and a delivered file.

  1. Segment resolution

    Planned

    Approved script segments are resolved with their pronunciation guidance and voice assignment.

  2. Speech synthesis

    Planned

    Each segment renders through a provider adapter negotiated against the required capabilities.

  3. Bed and effect placement

    Planned

    Music beds and effects are placed against segment boundaries.

  4. Mixing

    Planned

    Ducking and transitions are applied across the assembled timeline.

  5. Loudness normalization

    Planned

    The mix is normalized to the target programme loudness.

  6. Metadata and chapters

    Planned

    Canonical chapter markers and episode metadata are embedded.

  7. Preview and export

    Planned

    An editor previews the result before delivery formats are written.

Speech and provider boundaries

Text-to-speech providers differ in what they accept. The adapter boundary is where that difference is absorbed.

Providers vary in the markup they support, the voices they expose, the length of text they accept in one request, and whether they can render a fragment deterministically twice. Rather than assume a lowest common denominator, the design negotiates capabilities: a request declares what it needs, and an adapter either satisfies it or declines.

Speech synthesis sits behind a port like every other external service, so a provider change is an adapter change and does not reach the domain.

ADR 011: TTS capability negotiation (opens the Episodic repository)

Pronunciation control

A shared repository of how words should be said, rather than corrections scattered through scripts.

Names, domain jargon, and internal terminology are the recurring problem: a synthesizer will say them confidently and wrongly, the same way every episode. The pronunciation repository holds guidance once, and speech markup applies it at render time.

Chiltern reads the same repository when deciding whether a term in a draft still lacks guidance, which is why the check and the fix share one source rather than drifting apart.

ADR 012: Pronunciation repository (opens the Episodic repository)

Mixing and loudness

Ducking, transitions, and a single programme loudness target.

The mixing engine places beds and effects against segment boundaries, ducks them under speech, and handles transitions between segments. The output target is -16 LUFS integrated programme loudness, which is the conventional level for spoken-word podcast delivery.

Preview, metadata, and export

The last stage before an episode leaves the system.

Chapter markers generated against the script carry integer-only ISO 8601-style start times in PT#H#M#S form and are validated against the segment transitions they claim to mark. Because the canonical document already holds them, audio publishing projects them into whatever chapter format a player expects rather than regenerating them.

Preview and export jobs are roadmap step 3.6. When audio samples exist and belong to the project, they will appear here with a transcript, duration, format, and a descriptive label - under explicit controls, and never playing on their own.

Back to quality and review · Audio tasks in the roadmap