Fix Pronunciation and Pacing in AI Narration: Documented Controls and Methods

Fix a mispronounced word with the tool's pronunciation feature if it has one, and fall back to respelling the word; fix pacing by changing punctuation and splitting long sentences, then by using a pause control if the tool documents one. ElevenLabs, Google Play, Spotify's creation tools and Murf all document pronunciation controls, with limits that decide which fix you can use. This page covers the documented controls and then the methods. Finding the errors in the first place is the job of the proofing checklist.

Which controls each tool documents

The table lists what each vendor's own pages described on 2026-11-08. "Not stated" means the pages read were silent, not that the feature is absent. Each tool has its own section below.

ElevenLabs pronunciation dictionaries

ElevenLabs documents pronunciation dictionaries as XML .pls files, as of 2026-11-08. Each lexeme maps a word to either a phoneme, written in IPA or CMU, or an alias. The limits matter:

• Phoneme tags work only on eleven_v4, eleven_v4_turbo, eleven_flash_v2 and eleven_v3, per the docs. Other models ignore them, so use an alias there.
• The Studio page disagrees. ElevenLabs' Studio docs say phoneme tags are only compatible with Eleven Flash v2. Until the two pages agree, an alias is the safer starting point in Studio. If you use a phoneme, preview it on the model you will export with.
• Matching is case sensitive, so "Nate" and "nate" are different entries.
• Only the first match counts. ElevenLabs' best-practices guide says the dictionary is checked from start to end and only the first replacement is used, so a duplicate entry further down does nothing.
• Phoneme tags work on single words. A first and last name each need their own entry, per the same guide.
• The docs state no entry or file-size limits. That is silence, not a guarantee.
• Where it lives: the dictionary page we read sits in the agent-platform section of the docs, but the Studio docs describe a Pronunciations Editor with alias or phoneme rules, and the best-practices guide says dictionaries work in Studio. An older help article for the editor could not be read (the help site returned an access error).

An alias entry looks like this, inside the lexicon wrapper shown in ElevenLabs' docs. The name and respelling are our illustration:

<lexeme>
  <grapheme>Siobhan</grapheme>
  <alias>shi-VAWN</alias>
</lexeme>

For books, ElevenLabs' audiobooks page says pronunciation entries are pre-filled from your manuscript and applied book-wide. That is a vendor claim and the page gives no detail. Studio's documented limits include 5,000 characters per paragraph, which also caps how much you can reword into one block. The text-to-speech overview does not mention SSML or pronunciation dictionaries; the best-practices guide is where they are covered. For plan costs see the ElevenLabs review.

ElevenLabs pacing controls

ElevenLabs documents three pacing controls in Studio, and which one works depends on the model.

• Breaks. The Insert break button adds a tag, 1 second by default and 3 seconds at most. The best-practices guide says Eleven v3 and v4 do not support SSML break tags and that too many in one generation can cause instability. On those models it points to punctuation, text structure and audio tags instead.
• Speed. A voice setting from 0.7 to 1.2. The docs warn that extreme values may affect quality.
• Stability. A slider that sets how much each generation varies. ElevenLabs says lower values widen the emotional range and can produce odd, overly random performances, and that higher values can turn monotonous.

Two workflow details affect cost. Adding a dictionary marks the affected pieces of a project as unconverted, so they need generating again. And generating consumes credits, though Studio's docs say that when the button reads Regenerate you can regenerate twice without credits if you do not change the voice or text. The pricing page says only that a limited number of free regenerations may be available. A voice-setting change applies to every use of that voice unless you enable Override settings on the selection, and Lock paragraph protects a paragraph you are happy with.

Google Play's pronunciation check and pauses

Google's auto-narration editor documents the most complete word-level review of the routes we read, as of 2026-11-08. Its Pronunciation check groups words into edits, person names, organizations, locations, spelled-out words and spelled-out numbers. Pressing Play beside a word plays every occurrence in the book, which is a quick way to hear one name across chapters.

To change a word, right-click it, choose Edit Pronunciation, then pick one of the following:

1. An alternative pronunciation, if one is offered.
2. A phonetic spelling. Google's example replaces "1060" with "one thousand and sixty".
3. IPA text.
4. Speaking the word into your microphone. Google notes this is not available for some languages.

Google says the check can also find pronunciations for words spelled the same but pronounced differently. For pauses, choose the word you want the pause before and click Add a pause under Default narrator. Google suggests pauses for scene breaks and word emphasis. Voice and speed are set under Default Narrator, with no speed range stated. See Google Play auto-narration for the program's rules.

Spotify's pronunciation review

Spotify's Audiobook Creation Tools detect terms that may need review when you upload your EPUB, and you correct them with a phonetic spelling, per its help page as read on 2026-11-08. Spotify's example is "Bee-YON-say" for Beyoncé. You can listen to each term, skip it, or add names and acronyms by hand. Saved pronunciations carry into chapter previews and the final audiobook.

The limit to plan around is regeneration. Previews can be regenerated as often as needed before publishing, but Spotify's preview page says pronunciations lock for that version at Submit and Publish, while its editing page says you can update them afterwards by creating a new version, which uses one of the 4 generations. Each project allows up to 4 generations, including those made after publishing. If the first publish uses one generation, which is our reading, three remain for corrections. Spotify's pages document no pause or speed control. See Spotify's creation tools.

Murf pronunciation and pacing options

Murf's help page lists smart suggestions, custom IPA, alternative spelling and a per-word locale change, with the locale change noted for Gen2 voices. Murf's help page says the Pronunciation Library is Enterprise only; we did not read how other plans store corrections. Murf's audiobooks page names speed, pitch, emphasis and emotion controls and mentions adding pauses in passing, without values. Murf's pricing page could not be read on 2026-11-08, so check which plan you need. See the Murf review.

Three details from the help page to note:

• Paste IPA without slashes. Murf says slashes copied from a dictionary can trigger an "Invalid Character Found" error, so /wund/ goes in as wund.
• Alternative spelling handles words with two readings. Its examples are "livv" for the verb live, "laiv" for the adjective, and "em-bee-see" for MBC.
• Scope is a choice. The locale steps offer Apply for the selected block or Apply to Project for the whole project. If a word repeats and you are not on Enterprise, check what your plan offers before you correct it block by block.

Amazon Polly, for authors comfortable with code

Amazon Polly exposes pronunciation through its API rather than an editor. It takes lexicons in the W3C Pronunciation Lexicon format, the same family of file ElevenLabs uses, and they are stored per AWS region. Its SSML table, as read on 2026-11-08, lists break, phoneme and sub with full availability on neural and long-form voices, prosody as partial, and emphasis as not available on neural, long-form or generative voices. Polly returns an error for unsupported tags in standard, neural or long-form formats, so one bad tag can stop a request.

What we could not confirm for other routes

For the rest, no usable correction documentation was read, and the table says so plainly.

The INaudio limit is the one that changes your plan. Its FAQ says not to modify, convert, rename, extract or replace files in the package, so any fix has to happen in the generating tool. See Voices by INaudio.

Which fix to try first

Start with the smallest fix that could work, change one thing at a time, and listen before moving on. This order is our reasoning from the documented controls, not a measured ranking.

One trap: a dictionary entry replaces the word wherever it appears. ElevenLabs says "whenever one of these words is encountered", so a book-wide alias for "wound" would force one reading on both meanings.

Pre-flight bench

Check a passage before you spend credits

Paste a chapter or a tricky passage. The bench flags the things this guide says to check, and can read it aloud in your device's built-in voice so you can catch names and numbers that come out wrong.

Methods that need no dictionary

These are reasoned methods, not measured results. Try each on a single paragraph before applying it book-wide.

1. Respell the word. If a name is wrong, replace it in the narration copy with a phonetic spelling, such as "shi-VAWN" for "Siobhan". This is what an alias does inside a dictionary. Respelling in the source text changes your ebook too on routes that can read the live ebook (KDP requires a live eBook; Google's pages also allow a newly uploaded .epub), so keep a separate narration copy where you can. The preparation page covers copies. Hyphens and a capitalized stressed syllable are the pattern in Spotify's "Bee-YON-say", and ElevenLabs also suggests capitals, dashes and apostrophes as respelling tricks.
2. Change the punctuation for pacing. We would expect, though no source measures it, that commas and periods shorten or lengthen a pause. Add a comma where a voice rushes a clause and replace a comma splice with a full stop where it runs on. Listen after each change, because tools differ.
3. Split long sentences. We would expect, though no source measures it, that a sentence with several clauses is where some voices lose their place. Break it at a natural pause, keeping your meaning.
4. Use paragraph breaks for larger pauses. We would expect, though no source measures it, that a new paragraph or a scene-break line is likelier to produce a longer gap than a comma.
5. Rewrite instead of fighting. If a word will not come out right, a synonym costs less than ten retakes.
6. Record every fix. A running list of respellings makes the same name consistent across chapters.
7. Check the fix in its sentence. A word that sounds right alone may not survive its neighbors, so preview the whole sentence before you accept it.

Illustrative respellings, written by us and not checked against any voice:

Check each fix against the error log from your proof-listen. If you are still choosing a tool, a same-chapter comparison can include a round of these corrections, and the AI audiobooks hub has the next guide.

Last verified 2026-11-08. Rules and prices change: check the linked primary page before you rely on this.