When AI text to speech sounds bad, people blame the voice. Most of the time the voice is fine and the input is broken. The same handful of mistakes show up across every team that starts producing synthetic narration, and each one has a clear cause, a measurable cost, and a corrective practice you can adopt in minutes.
This is a field guide to those failure modes. We will name seven, explain why each happens, what it costs you, and exactly what to do instead. If you understand the pipeline behind these tools, the causes will make sense; if you do not yet, What Actually Happens Between Your Text and the Voice is worth a read first. The point here is correction, not theory.
Mistake 1: Feeding the Model Raw, Unnormalized Text
The single most common error is pasting text exactly as written and assuming the engine will interpret it like a human. It will not. The model reads what is there, so Dr. becomes "doctor" when you meant "drive," 2024 becomes "two thousand twenty-four" when you meant a year, and $1.5M becomes a mess.
Why it happens: People think of TTS as reading, but the normalization stage is literal. The cost: confident-sounding mistakes that survive into the final cut. The fix: spell out ambiguous items in your script, or write the number the way it should sound. Clean input is the cheapest quality lever you have.
Mistake 2: Auditioning Voices on the Demo Sentence
Every platform has a polished demo line that makes every voice sound great. Teams pick a voice based on that line, then discover it grates over a five-minute script.
Why it happens: the demo is engineered to flatter. The cost: you commit to a voice, render hours of content, and only then hear the fatigue. The fix: always audition with a representative chunk of your actual script, ideally one that includes your hardest words and a range of sentence lengths. Listen for fatigue, not first impressions. The Repeatable Workflow for Producing Clean AI Narration bakes this into the process.
Mistake 3: Ignoring Pronunciation of Names and Brands
Proper nouns are where TTS stumbles, because names often are not in the pronunciation dictionary and the model has to guess. A mispronounced product or client name in an otherwise flawless render reads as carelessness.
Why it happens: the grapheme-to-phoneme model guesses for unknown words. The cost: credibility, especially in client-facing audio. The fix: build a custom lexicon. Define the phonetic spelling of every brand, product, and recurring proper noun once, then reuse it across projects. This is the highest-leverage fix on this list.
Mistake 4: Over-Tuning Pitch and Rate
Beginners discover the pitch and rate sliders and crank them. The result sounds artificial, chipmunked, or unnaturally slow.
Why it happens: the controls feel powerful, so people overuse them. The cost: a voice that screams "synthetic." The fix: treat these controls as seasoning. A slight reduction in rate for comprehension is usually all you need. Large pitch shifts almost always sound worse than the default.
The punctuation alternative
Before reaching for pitch and rate, edit your punctuation. A comma, a period, or splitting a sentence in two reshapes pacing more naturally than any slider, because it works with the acoustic model instead of against it.
Mistake 5: Rendering the Whole Thing on the First Try
Someone pastes an hour-long script, hits generate, waits, and discovers a pronunciation error in minute three that poisons the whole file.
Why it happens: impatience and false confidence. The cost: wasted render budget and time, plus the temptation to ship the flawed version. The fix: always generate a short test paragraph first, one that includes your trickiest words. Validate, then commit to the full render. Catching problems on a paragraph costs seconds.
Mistake 6: Editing the Audio Instead of the Source
When a word sounds wrong, the instinct is to open an audio editor and patch it. This is almost always a mistake.
Why it happens: audio editing feels more direct. The cost: the fix is not reproducible, so the moment you need to re-render, the problem returns. Audio editing also cannot truly fix pronunciation. The fix: correct the text or the lexicon, then re-render. Text-level fixes are reproducible and survive future changes. Treat the script and settings as the source of truth.
Mistake 7: Cloning Voices Without Consent or Disclosure
The most serious mistake is not technical. Voice cloning can produce a convincing synthetic version of a real person, and teams sometimes do it without explicit permission or without disclosing that the audio is synthetic.
Why it happens: the feature is easy and the governance is an afterthought. The cost: legal exposure, broken trust, and reputational damage that no render quality can offset. The fix: treat voice cloning as a consent-and-disclosure question first and a feature second. Get written permission, disclose synthetic audio where it matters, and document your policy. The technology is neutral; your process is not.
How These Mistakes Compound
Individually, each mistake is annoying. Together, they reinforce each other into a workflow that feels cursed. Unnormalized text (Mistake 1) makes the model mispronounce things, which tempts you to patch the audio (Mistake 6), which then breaks the moment you re-render after a script change. Skipping the test render (Mistake 5) means you discover all of this at the worst possible time, on the full file, near a deadline.
The pattern behind every one of these is the same: trying to fix downstream what should have been controlled upstream. The corrective mindset is to push every decision as early in the pipeline as it will go. Resolve ambiguity in the script, not the settings. Fix pronunciation in the lexicon, not the audio. Validate on a paragraph, not the full render. When you adopt that upstream-first discipline, most of these mistakes stop being possible rather than just rare. The The SHIP Model for Reliable AI Voice Production is one way to make that discipline a habit your whole team shares.
Frequently Asked Questions
Which mistake is the most expensive to make?
Cloning without consent, by a wide margin, because the cost is legal and reputational rather than just a wasted render. Among quality mistakes, rendering a full file before testing wastes the most time and budget. Both are entirely preventable with a process change.
How do I know if a problem is normalization or the voice itself?
If the model says the wrong word, that is normalization or pronunciation, fixable in the text or lexicon. If the word is correct but the delivery sounds robotic or flat, that is the voice and prosody, addressed through voice choice and punctuation. Naming the stage tells you where to fix it.
Is it ever fine to edit the audio directly?
For final polish like volume normalization or trimming silence, yes. For correcting pronunciation or pacing, no. Those belong in the text and settings so they survive a re-render. Reserve audio editing for things the source text genuinely cannot control.
Why do my custom pronunciations sometimes still sound wrong?
Pronunciation can shift based on surrounding words, so a lexicon entry that works in isolation may need testing in context. Re-test the fix inside the actual sentence rather than alone. Some platforms also scope lexicons per language, so confirm you defined the term for the right voice.
How often should I update my lexicon?
Add an entry every time you hit a new name, brand, or acronym the engine gets wrong. Over a few projects this becomes a reusable asset that eliminates most pronunciation work. Treat it as living documentation rather than a one-time setup.
Key Takeaways
- Most "bad voice" problems are actually bad input; clean and normalize your text first.
- Audition voices on your real script, not the flattering demo line, and listen for fatigue.
- Build and reuse a custom lexicon for names and brands; it is the highest-leverage fix.
- Use pitch and rate sparingly; edit punctuation before reaching for sliders.
- Always test on a short paragraph before a full render, and fix the source, not the audio.
- Voice cloning is a consent-and-disclosure question first; never clone without permission.