Skip to content

Overview & Architecture

AudioBard is designed with a modular, local-first architecture that prioritizes privacy, deterministic output, and audio quality.

The 6-Stage Pipeline

When you process a book through AudioBard (either via the Desktop GUI or the CLI generate command), it passes through six decoupled stages:

flowchart TD
    A["Book Input (.epub / .txt)"] --> B["Parser & Chapter Splitter"]
    B --> C["Character Extraction (LLM)"]
    C --> D["Tone-Aware Voice Casting (Deterministic Hash)"]
    D --> E["Dialogue Attribution & Emotion Tagging (LLM)"]
    E --> F["Async Neural TTS Synthesis (Piper / Edge)"]
    F --> G["Audio Normalization & Assembly (FFmpeg / pydub)"]
    G --> H["Final Audiobook (.mp3 / .m4b)"]

Stage-by-Stage Breakdown

1. Ingestion & Pre-Processing

  • Accepts .epub and .txt files.
  • Strips Project Gutenberg disclaimers, headers, and license footers.
  • Preserves paragraph formatting and dialog punctuation (—, «», ").

2. Character Roster Extraction

  • Feeds the opening ~5,000 words of the text into the configured LLM.
  • Returns a structured JSON roster of characters, including:
  • canonical_id (e.g., Character_Elizabeth_Bennet)
  • aliases (e.g., ["Lizzy", "Miss Bennet", "Elizabeth"])
  • gender_hint and age_hint
  • tone / temperament descriptor
  • Validated via Pydantic schemas.

3. Tone-Aware Voice Casting

  • Matches extracted characters against the available voice pool for the target locale (en_US, es_MX, es_ES, es_CO).
  • Filters by gender and age hints.
  • Computes semantic similarity between character tone and voice profile tags.
  • Uses a deterministic hash tie-breaker so subsequent runs of the same book assign identical voices.

4. Sliding-Window Dialogue Attribution

  • Splits the manuscript into ~1,500-word chunks with a 5-paragraph overlapping window.
  • The LLM attributes each sentence to a character or "Narrator".
  • Assigns an emotion tag (neutral, happy, sad, angry, whisper, surprised, sarcastic).

5. Async Neural TTS Synthesis

  • Synthesizes speech per line with prosody adjustments (speed, pitch, pauses) based on emotion.
  • Multi-layer caching: In-memory LRU + SQLite persistent disk cache (~/.cache/audiobard/tts/).
  • Only uncached lines call the TTS engine.

6. Assembly & Mastering

  • Normalizes volume across distinct voices to prevent loudness jumping.
  • Injects configurable silence gaps between narrator descriptions and spoken lines.
  • Exports to standalone .mp3 or chapter-tagged .m4b.