StoryWriter

Turns an audio track into a short story whose structure is inferred from the sound — not a one-dimensional "loud = happy" mapping, but multiple characters and intersecting arcs read from the music's layers, contrasts, peaks and crossings.

The pipeline

upload mp3
  → extract_arcs_hpss.py    audio analysis (one NumPy STFT → 7 feature arcs over time)
  → server.py               downsamples features for the live graph
  → story_generator.py      THE MODEL DRIVES:
        1. reads the audio's structure (layers, contrasts, peaks, crossings) and
           infers the SEQUENCE OF EMOTIONAL STATES it moves through — not
           one-dimensional
        2. emits a POINTS line: a JSON array of 5–9 emotional waypoints
           (valence/arousal/dominance, 1–9) for how the piece FEELS, moment to moment
        3. WRITES one coherent short story that traces that emotional arc — a single
           drafting pass — optionally studying a published opening page for voice

The emotional arc comes entirely from the model reading the audio. (Earlier versions queried a MongoDB word/phrase lexicon per waypoint; that path has been removed.)

The generator streams newline-delimited JSON events on stdout (status, reasoning, lookup, story, story_reset, final, error); server.py turns those into the job state that /status serves, so the browser shows the audio graph filling in, the model's reasoning, each emotional waypoint as a dot, and the story typing out.

Files

File	Role
index.html	The UI (React, single file). Served at / and proxied to /storywriter.
server.py	Backend on :8000. Upload → extract → generate; serves /status.
story_generator.py	The model-driven story engine.
extract_arcs_hpss.py	Pure-NumPy audio-feature extractor → *_arcs.csv.
build_openings.py	Loads public-domain opening pages into MongoDB (style only, optional).
run-storywriter.ps1	One command: check deps/Ollama, start the stack, open the browser.
restart-storywriter.ps1	Stop/relaunch just the backend on :8000.
install-storywriter-task.ps1	Register the boot auto-start task.

Prerequisites

• Ollama running a model (e.g. qwen2.5:7b): ollama list.
• Python 3.12 with pymongo, numpy, imageio-ffmpeg.
• MongoDB is optional — used only for the opening-page style exemplars (build_openings.py). Generation works fine without it.

Run

powershell -ExecutionPolicy Bypass -File .\run-storywriter.ps1

That checks dependencies, starts the backend (:8000) and gateway (:80), and opens http://localhost/storywriter. Standalone (no gateway): http://localhost:8000.

Config via env vars: STORYWRITER_OLLAMA, STORYWRITER_MONGO_URI, STORYWRITER_OPENINGS_DB, STORYWRITER_OPENINGS_COLL, STORYWRITER_TOKEN.