How VoxAnvil fixes numbers, abbreviations, and phonetics before narration

Why a novel is harder to narrate than a paragraph

Speech engines read what they are given. A novel is full of strings that are easy for a human and awkward in raw text. A price such as $2,500 can be read as digits and punctuation. Dr. can be read as the letters D and R, or as Doctor, or with the period spoken. St. can mean saint or street. A date such as 11/05/2026 can be a month-first date or a day-first date, and the wrong choice is confident rather than hesitant. Invented names have no dictionary entry at all.

Regenerating the chapter does not fix that class of problem if the text stays the same. The engine is likely to make the same decision. The useful work happens in the manuscript: decide the spoken form, store it, and send that form. VoxAnvil calls this preparation layer the Manuscript Intelligence Engine. In the product it is not a single magic rewrite of the file. It is a scan that proposes replacements, a project dictionary that applies the replacements you keep, and a review report that looks for production risks before you export.

The public site includes a preflight demonstration with sample flags for currency, titles such as Dr. and Mr., a contraction, an em dash, and a numeric date. That demonstration shows the kinds of issues the engine is built to catch. The lasting control, the one that changes the audio, is the rule you save on the project. A suggestion you never save does not rewrite the chapter by itself.

What the preflight scan returns

The preflight scan takes manuscript text and asks for a structured list of narration risks. Each flag has a type, the original text, a suggested spoken replacement, a short bit of surrounding context, and an approximate position. The types the scanner is instructed to use are number, abbreviation, contraction, name, punctuation, currency, date, formatting, and other. The instructions focus on writing numbers and currency out in words, expanding abbreviations such as Dr., Mr., St., and Ave., handling contractions, reading dates as spoken dates, and noticing em dashes, special characters, URLs, and acronyms.

The scan accepts a long paste, and the analysis itself looks at the first portion of that paste, about 8,000 characters. A full novel will not be exhaustively flagged by one call. Scan the chapter in front of you, or work from the pronunciation report after you have saved the names you already know. Treat the suggested text as a draft reading. You know whether 1998 is a year or a count. You know whether St. Paul is the city or a street. The scanner can propose a fluent phrase. You accept the meaning.

Starter’s plan description calls this manuscript preflight for number and abbreviation fixes. Pro’s plan description calls it the full preflight suite and pairs it with the narration review. Those are the labels on the pricing page. In practice, the scan’s job is to show you the phrase and a spoken alternative. Pronunciation Forge and your own edit of the chapter are how that alternative gets into the audio.

Pronunciation Forge, the project dictionary

Pronunciation Forge is a dictionary on the project. Each rule has the phrase as written, the way it should be spoken, a type, and a switch for case sensitivity. Types are word, name, acronym, and phrase. Free plans see the tool and are asked to upgrade. Starter and Pro can save, update, and delete rules. A rule belongs to that project and that account. It does not silently become a global dictionary for every book on the platform.

At generation time the studio applies enabled rules to the chapter before the ElevenLabs request. Longer phrases are replaced first. A rule that starts and ends with a word character is matched on word boundaries, so a short name does not fire inside a longer word. Case-sensitive rules leave other casings alone, which is how you can teach an acronym without also rewriting a common word that happens to use the same letters in lowercase. You can preview the spoken form with a short synthesis of that phrase before you add the rule.

This is the phonetic fix authors actually feel. If a character is named Aeltharyn, you store the syllables you want, such as a respelling that a narrator will say correctly. If an acronym should be spelled letter by letter, the spoken form is those letters separated. If it should be a word, the spoken form is the word. The manuscript can keep the spelling your readers see in print. The audio receives the spoken form. That separation is the point of the dictionary.

Write the spoken form as words or a simple respelling, not as a phonetic alphabet the voice has not been shown. The preview button is the test. If the preview is wrong, change the spoken form before you generate the chapter. A bad rule is applied everywhere the phrase appears, which is powerful when it is right and loud when it is wrong.

Narration QA on the Pro plan

Narration QA is the Pro review that runs against the manuscript after you have text in the project. It refuses to run on an empty project and asks you to add manuscript text first. The deterministic part of the report does not depend on a model being in a good mood. It looks for an unclosed square bracket. It looks for bracket tags outside the supported set of expression, pause, and sound-effect tags. It looks for a sentence that appears more than once, which is often a paste error. It looks for a very long sentence, for a heavy density of ellipses, for acronyms of two or more capitals that do not already have a pronunciation rule, and for a long all-caps line that a voice may shout.

Chapter roman numerals are not treated as mystery acronyms. The check is aimed at initials and abbreviations you may want spoken a specific way. The suggestion on an unmapped acronym is narrow: add a pronunciation rule only for the ones that need a specific reading. You do not have to dictionary every capital letter in the book.

A second pass can add up to a handful of editorial notes about pronunciation, pacing, accidental repetition, dialogue, tone tags, or formatting. If that pass is unavailable, the deterministic findings still come back. The combined list is scored from 0 to 100, with critical issues weighted more heavily than warnings, and warnings more heavily than notes. The summary tells you whether to fix a critical issue before narration or whether the remaining items are reviews. Ready to export, in this report, means the score and the issue list. It does not mean a human mastering engineer has signed the file.

A practical order of work

Do the dictionary before the first full generation of a chapter. Read the opening pages and list proper names, place names, coined words, and acronyms. Add those rules and preview the ones you are unsure about. Then run a preflight scan on that chapter’s text and look at numbers, money, dates, and abbreviations. Put the readings you agree with into the dictionary or into the chapter text. On Pro, run Narration QA and clear unclosed tags and repeated lines before you spend a generation on them.

Generate one chapter and listen with the page in front of you. When a word is wrong, add or edit the rule and regenerate that chapter. Rules apply on the next generation because they are substituted at request time. They do not rewrite audio you have already stored. If a chapter was generated before the rule existed, produce it again.

Keep retailer pronunciation questions in the same pass. A name you will say in an opening credit should match the name in chapter one. The distribution checklist can remind you that credits files exist. The dictionary is what makes the name in those files match the name in the story. When the chapter sounds like the book you wrote, continue to the next one.

Frequently asked questions

Related reading

• The Manuscript Intelligence Preflight Method
• How AI audiobook narration works
• What is VoxAnvil?
• VoxAnvil and a DIY ElevenLabs workflow
• VoxAnvil glossary