IN-HOUSE PROJECT / CASE STUDY / BUILT BY THE SILO FORGE

An in-house pipeline,
built end-to-end.

ScriptCast is an in-house project by The Silo Forge: a local-first pipeline that compiles declarative YAML scripts into mastered audio and video files (FLAC, MP3, M4B, MP4), using a local Python inference microservice and an automated FFmpeg DSP pipeline.

The empty house.yaml production script
name: The empty house settings: preRollMs: 12000 postRollMs: 9500 metadata: title: The Empty House author: Algernon Blackwood anthology: The Empty House and Other Ghost Stories year: "1906" description: "Dramatization of Algernon Blackwood's 1906 short story." coverImage: cover.png genre: Supernatural Horror mastering: audioProfile: EBU-R128 # Target -23.0 LUFS, -1.0 dBTP, LRA 20.0 videoProfile: YouTube # 1080p 24fps (h264_nvenc / libx264), AAC 384k limiterEnabled: true limiterAttackMs: 5 limiterReleaseMs: 50 loudnessNormalizationEnabled: true characters: - id: narrator - id: shorthouse - id: aunt_julia dialogue: - speaker: narrator instruction: "Low and unhurried, quietly ominous" pauseAfterMs: 2200 - speaker: shorthouse instruction: "Forced bravado, in a half voice" environment: Hall reverbMix: 0.42 pauseAfterMs: 2500 musicCues: - number: 1 fromLine: 1 toLine: 5 file: mus_1_overture_thesis_of_evil.flac duration: 108 duckUnderDialogue: true duckAmountDb: -6 offsetMs: -12000 anchor: Start ambientCues: - number: 5 fromLine: 54 toLine: 76 file: amb_5_empty_house_hall.flac duckUnderDialogue: true duckAmountDb: -3 loopCrossfadeMs: 2000 images: - file: img-cover.jpg filters: - type: noise strength: 40 - type: fadeIn durationMs: 2000

A single schema encapsulates character voice profiles, line acoustics, continuous looped beds, sound design cues, deterministic ducking policies, and target mastering profiles.

From script to mastered files.

Processing runs locally from start to finish. Audio files are validated, synthesized, aligned, normalized, and packaged without leaving the workstation.

1

Preflight Checks

Scans the physical project hierarchy, validates referenced audio stems across localized category directories, queries the inference service health endpoint, and verifies FFmpeg/FFprobe system binaries.

2

Local Synthesis & Queue

Dispatches text, reference audio, and generation parameters to the inference service over HTTP. Supports modular voice and audio engines (Breeze for voice synthesis and voice design, Stable Audio for cues) with individual item regeneration.

3

Stem Normalization & Cache

Calibrates individual audio files to category reference levels (Dialogue: -20.0 LUFS, Music: -32.0 LUFS, Ambience: -34.0 LUFS, SFX: -6.0 dBFS/dBTP). Measurements are stored in lufs.yaml to skip re-measuring unchanged files.

4

Sample-Accurate Timeline & Frame-Evaluated Ducking

Calculates timeline coordinates from physical audio durations and punctuation-based prosody. Generates an FFmpeg volume expression (volume=...:eval=frame) anchored to dialogue boundaries with grouping gaps and edge crossfades to prevent pumping artifacts.

5

Two-Pass Mastering, Tagging & Video Delivery

Applies two-pass loudness normalization (loudnorm) and peak limiting. Injects tags and chapter timecodes into FLAC, MP3 (320 kbps), and M4B (AAC 256 kbps) containers via ATL. Extracts promotional MP3 samplers and renders 1080p MP4 video using NVENC or software encoding with modular filter chains.

Signal processing and mastering defined in script.

Audio properties are declared directly in text or edited through dedicated application panels. The pipeline builds and executes FFmpeg filter chains automatically.

Avalonia Desktop Studio & SoundFlow In-Process Playback

Cross-platform desktop interface with virtualized asset lists, search and ungenerated filters, and in-process audio preview via SoundFlow (MiniAudio device backend), avoiding external process spawning and buffer underruns.

Multi-Language Sidecar Architecture

Maintain localized productions in English, European Spanish, or Latin American Spanish. Structural cues and timeline boundaries remain centralized in the master script while transcripts and audio assets resolve via language sidecars (e.g. script.es-es.yaml).

Asset Synthesis & Queue

Synthesize entire productions or enqueue individual dialogue takes, character voice profiles, sound effects, and ambient beds into an asynchronous task queue with real-time execution tracking.

Production Profiles

Configurable mastering presets targeting industry standards: EBU-R128 (-23.0 LUFS, -1.0 dBTP), ATSC-A/85 (-24.0 LUFS, -2.0 dBTP), Podcast-AES (-16.0 LUFS, -1.0 dBTP), and Audible-ACX (-20.0 LUFS, -3.0 dBTP).

Spatial & Environmental Acoustic Filters

Calculates distance attenuation with air absorption modeling, occlusion filtering, dynamic timeline pan sweeps (from start to end coordinates), stereo widening, and parametric acoustic presets (Room, Hall, Cave, Telephone).

Equal-Power Loop Tiling

Continuous ambient beds and musical drones are pre-processed into seamless loop tiles using a constant-power quarter-sine (qsin acrossfade) window to maintain constant signal energy across boundaries.

Video Packaging & Sampler Generation

Muxes mastered audio streams with visual cues into YouTube-compliant 1080p MP4 files using NVENC hardware encoding (fallback to libx264) with Ken Burns zoom-pan drift, grain, vignette, and opacity fades. Procedurally extracts and crossfades promotional audio samplers.

Pilot: The Empty House.

Production run exercising batch orchestration and DSP on a complete screenplay.

Integration Run

The Empty House

Algernon Blackwood · 1906 · Multi-Cast Drama · Mastering: EBU-R128 / YouTube MP4

Integration run across a 176-dialogue screenplay, 15 looped ambient beds, 21 music cues, and 20 SFX cues using acoustic environment presets (Room, Hall, Cave, Telephone).

The pipeline exercises batched inference through the local microservice, ducking matrix generation, stem normalization, two-pass mastering, metadata tagging, and video encoding into distribution files.

Dialogue Lines
176 Lines (Multi-Cast)
Musical Cues
21 Scored Cues
Ambient Beds
15 Looped Beds
Sound Effects
20 Foley Cues
Loudness Target
-23.0 LUFS / -1.0 dBTP
Deliverables
FLAC, MP3, M4B, MP4 Video

Implementation notes.

What has been built and what is being validated. This is a case study of a working system, not a product roadmap.

Built

Avalonia Desktop Studio & YAML Parser

Cross-platform GUI with virtualized item lists, script lifecycle persistence, in-process SoundFlow MiniAudio playback, and multi-language editing dialogs.

Built

Local Inference Client

HTTP request dispatch and execution tracking against a local Python inference microservice (Breeze, Stable Audio).

Built

DSP, Ducking & Mastering

Timeline-anchored frame-evaluated ducking envelopes, continuous pan sweeps, equal-power loop tiling, and two-pass loudness mastering (EBU-R128, ATSC-A/85, Podcast-AES, Audible-ACX).

Built

Packaging & Video Rendering

Master FLAC export, ID3v2 MP3, chaptered M4B with embedded artwork via ATL, promotional sampler extraction, and NVENC/libx264 1080p MP4 video delivery.

Built

Unit Tests

Unit test suite covering core domain contracts, FFmpeg filter generation, video packaging and inference client adapters, and ViewModels using MockFileSystem and NSubstitute.

In progress

Pilot Production Execution ("The Empty House")

End-to-end batch synthesis and master file export of the complete 176-line screenplay.

Need something like this for your company — desktop systems, hardware integration, offline-first applications, or local processing pipelines? I take direct B2B contracts.

See services and engagement terms →