Pronunciation Dictionaries for Synthetic Audiobooks: Handling Accents and Fantasy Names

High-volume fiction publishers and indie authors extending a backlist into audio soon hit a familiar production barrier. Hiring human voice talent for a single 10-hour audiobook costs $2,000–$4,000 in finished hour rates ($200–$400 per finished hour). Across a twenty-book catalog, that represents $40,000 to $80,000 in upfront production capital—an investment that is nearly impossible to recoup through backlist sales alone.

Recent open-source text-to-speech engines like Kokoro TTS have upended this economic equation. Running locally on consumer hardware, these models synthesize natural, highly expressive human narration at zero incremental cost per hour. However, when publishing fantasy, science fiction, or historical romance, authors face a distinct technical challenge: phonetic drift.

Without explicit guidance, a neural voice model relies on default grapheme-to-phoneme (G2P) rules. A fantasy character named "Aeloria" or a sci-fi realm named "Xylos" might be pronounced three different ways across ten hours of audio—or mangled differently between Book 1 and Book 3. To maintain listener immersion and achieve professional AI audiobook pronunciation, publishers must implement a custom lexicon TTS dictionary. Here is how to construct phonetic dictionaries, manage regional accents, and lock down voice consistency across an entire series.

The Mechanics of Phonetic Drift in Synthetic Speech

Neural speech synthesis models do not read words like human voice actors. Instead, an internal G2P algorithm converts printed text characters into sequence representations of phonemes before generating raw audio waveforms.

Standard English G2P algorithms handle everyday vocabulary reliably. Problems emerge when the model encounters out-of-vocabulary (OOV) terms, proper nouns, and fictional terminology:

• Unpredictable Vowel Shifts: An invented name like "Caelum" may be parsed as "Kay-lum" in one chapter and "See-lum" in the next, depending on surrounding punctuation and sentence context.
• Mangled Compound Names: Fantasy surnames containing apostrophes or capital letters (e.g., "D'Artan" or "Kaelthas") confuse standard word-boundary tokenizers, causing awkward pauses or dropped syllables.
• Inconsistent Stress Assignment: Foreign or archaic loanwords often receive primary stress on the wrong syllable, turning a sinister villain's name into a comical mispronunciation.

Inconsistent pronunciations pull listeners out of the narrative and generate negative reviews on retail storefronts. As explored in our guide to mastering the soundstage and audio post-processing, meeting broadcast audio standards requires eliminating verbal artifacts before audio rendering begins.

Building a Custom Lexicon TTS Dictionary

A custom pronunciation dictionary (or lexicon file) overrides the speech engine's default G2P converter. When the synthesis pipeline encounters a word matching an entry in your dictionary, it bypasses default guessing and outputs your exact phonetic mapping every time.

Depending on your synthesis engine and workflow, phonetic dictionaries are defined using one of two primary formats:

1. Phonetic Respellings (Simple Lexicon Mappings)

For rapid setup, simple key-value mappings substitute complex proper nouns with phonetically unambiguous English respellings prior to synthesis:

• "Aeloria" → "Ay-LOR-ee-uh"
• "Xylos" → "ZY-loss"
• "Thalor" → "THAY-lore"

Simple respellings work well for basic proper nouns. However, because letter-to-sound rules in English remain inherently ambiguous, complex dialects or subtle vowel shifts benefit from standardized phonetic notation.

2. International Phonetic Alphabet (IPA) Mappings

Advanced TTS engines like Kokoro TTS natively support IPA phoneme sequences. IPA provides absolute precision over vowel quality, consonant articulation, and primary/secondary stress:

• "Caelum" → /ˈkae.lʊm/
• "Symphony" → /ˈsɪm.fə.ni/

By explicitly specifying IPA phoneme strings, you remove all ambiguity from the synthesis engine, ensuring identical pronunciation across every chapter and voice actor model.

Managing Regional Accents and Dialectical Shifts

Fiction frequently features characters with distinct regional accents—such as a Scottish-accented dwarf, a southern-accented detective, or a high-court aristocratic narrator. When building an audiobook, managing accents requires balancing vocal timbre with phonetic clarity.

Base Voice Models vs. Phonetic Shifts

In modern synthetic narration pipelines, regional accent is primarily determined by the base voice model's training dataset (e.g., selecting a British English or American Southern vocal profile). The custom lexicon should handle terms specific to that accent without overriding the natural inflection of the underlying voice model.

Preserving Dialect Spelling

When dialogue is written phonetically in the manuscript (e.g., "doin'" instead of "doing" or "lookin'" instead of "looking"), ensure your text preparation pass standardizes these contractions into your lexicon. Unchecked apostrophes in dialogue tags can cause synthesis engines to stutter or clip terminal consonants.

Series-Wide Voice Locking

For multi-volume book series, maintaining continuity across years of publishing requires locking down voice identifiers, synthesis speed parameters, and lexicon files in a centralized project configuration. As detailed in our guide on publishing a series without continuity breaking, storing your story bible alongside your audio parameters guarantees that Book 5 maintains identical character voices and terminology pronunciations as Book 1.

A Practical Workflow for Pronunciation Audits

To integrate custom dictionaries into a high-volume publishing operation without creating manual bottlenecks, follow a structured four-step audit pass:

1. Extract Out-of-Vocabulary Terms: Run an automated script or text scan over your completed manuscript to extract all capitalized proper nouns, invented terms, and foreign phrases not found in standard English dictionaries.
2. Define Canonical Mappings: Create a master dictionary file (pronunciation.json or .pls) mapping every extracted term to its preferred IPA string or phonetic respelling.
3. Render Isolated Sample Clips: Synthesize 5-second audio samples for every term in the dictionary. Audition the audio clips to verify that stress placement and vowel clarity sound natural.
4. Run a Dialogue Test Pass: Render a 2-minute sample passage containing heavy character dialogue and world-building terms. Verify that transitions between normal narrative prose and custom lexicon terms remain seamless.

Following this pre-render audit ensures your catalog satisfies distributor criteria. Reviewing the ultimate checklist for self-publishing an AI audiobook provides additional details on preparing file formats, chapter breaks, and metadata before distribution.

Automating Production in the Instawritr Pipeline

Manually editing lexicon files, running text cleanup passes, and configuring audio rendering parameters for every title can consume dozens of hours per book. To maintain a profitable backlist, high-volume publishers require an integrated pipeline that automates technical friction while keeping creative control in the author's hands.

The Instawritr pipeline automates manuscript creation, formatting, and audio synthesis within a single unified environment:

1. Synopsis: Expand your initial premise into a structured narrative framework.
2. Story Bible & Open-Loop Ledger: Instawritr compiles comprehensive character dossiers, world rules, and a dynamic open-loop ledger to track unresolved plot points across chapters.
3. Chapter-by-Chapter Drafting: The generation engine drafts rich, scene-focused prose one chapter at a time, eliminating context degradation and memory loss.
4. Automated QA Loop: An automated editorial pass evaluates the draft, eliminating passive phrasing, repetitive sentence starters, and continuity errors.
5. Cover Design & EPUB Compilation: Instawritr compiles an error-free, validated EPUB 3 container and generates a genre-appropriate cover using an integrated Google nano-banana image model.
6. Kokoro TTS Audiobook Synthesis: The clean EPUB passes directly into the local Kokoro TTS engine, applying your custom pronunciation lexicons, synthesizing chapterized audio, and mastering files to retail RMS and noise specs in a single automated step.

Because Instawritr is fully resumable, you can pause production at any checkpoint—whether to refine a character's IPA mapping in the story bible, adjust a scene outline, or edit prose—and resume drafting or audio rendering without losing progress or re-processing completed chapters.

For complete privacy and cost control, Instawritr can be executed fully local via llama.cpp, keeping your manuscripts, voice models, and generated assets entirely on your local machine with zero API fees. Once compiled, your books and audiobooks are staged for wide distribution across six major retail channels: ebooks to Apple Books, Google Play, Kobo, and Barnes & Noble; audiobooks to Spotify for Authors and InAudio. To review real proof, inspect a full book's production record on our teardown page, where you can also view its live listings across 51 Apple Books stores worldwide.

---

Ready to streamline your audiobook production stack and eliminate recurring subscription fees? Instawritr is available as a one-time purchase ($149 Pro / $249 Pro Plus), giving you a self-hosted publishing pipeline that supports bring-your-own AI API keys or fully local offline drafting via llama.cpp. Explore how it works to see our automated drafting and Kokoro TTS audio stack in action, or visit our pricing page to select the right tier for your publishing catalog.