One mispronounced brand name is enough to make a polished video feel cheap. The frustrating part is that it’s usually consistent: the same sentence breaks in the same wrong place, the same acronym gets spelled out every time.
This guide covers two things and skips the rest: making pauses sound like a person talking, and getting proper nouns right on every take. The first half uses standard SSML to explain the mechanics. The second half covers what still works when your tool only accepts plain text.
“It reads it wrong” is three different problems
People treat mispronunciation as one issue. It’s three, and they need different fixes.
| Layer | What you hear | Actual cause | Fix |
|---|---|---|---|
| Phrasing | A sentence split mid-thought | Too few commas, slashes and parentheses confusing the parser, sentences too long | Rework punctuation, split sentences, add break |
| Pronunciation | Acronyms, names, homographs wrong | Token missing from the lexicon; the model guesses from spelling | phoneme / sub, respell the word |
| Prosody | Every word correct, but flat or rushed | No variation in rate or pitch | prosody, split into separate takes, change voice |
Diagnose the layer first. No amount of prosody fixes a pronunciation error.
The six tags that solve most problems
The SSML spec is large. In practice, six tags cover roughly ninety percent of real issues. Engines support different subsets — the same tag can behave differently in Amazon Polly, Azure Speech, and Google Cloud TTS — so audition on the engine you’ll actually ship with.
| Tag | Purpose | Example | Watch out |
|---|---|---|---|
break |
Insert a pause | <break time="400ms"/> |
Past one second it sounds like a dropout |
sub |
Replace the spoken form | <sub alias="sass">SaaS</sub> |
Captions still show the original |
phoneme |
Specify phonemes | <phoneme alphabet="ipa" ph="ˈsoʊfə">Sopha</phoneme> |
Least portable; test per engine |
say-as |
Declare content type | <say-as interpret-as="characters">API</say-as> |
Most useful for numbers, dates, phone numbers |
prosody |
Adjust rate, pitch, volume | <prosody rate="95%">…</prosody> |
Keep changes small; ±10% is plenty |
emphasis |
Stress a word | <emphasis level="strong">free</emphasis> |
Overuse sounds like an ad read |
One rule that saves time: if sub works, don’t reach for phoneme. sub is portable across engines. phoneme varies, and every change forces you to re-test every segment.
Pauses: stop faking breaths with commas
The most common mistake is stacking punctuation — ,,, or ... — to manufacture a pause. The result is clutter on the page and dead-flat audio, because the parser collapses consecutive punctuation into one break.
Treat pauses as rhythm design instead, calibrated to meaning.
| Position | Duration | Why |
|---|---|---|
| Between words (natural breath) | 100–150 ms | Shorter smears words; longer stutters |
| Between clauses (comma) | 250–350 ms | Time to absorb half a thought |
| Between sentences (period) | 400–600 ms | A semantic unit ended |
| Paragraph or section change | 700 ms–1 s | Section feel; pair with a visual cut |
| Before a key line | 300–500 ms | Builds anticipation — the most effective emphasis there is |
| Between list items | 300 ms | Makes three items sound like three, not one run-on |
Same sentence, two versions:
<!-- Flat: all commas -->
Our platform cuts costs, improves efficiency, and gets teams live faster.
<!-- Paced: distinct levels -->
Our platform cuts costs.<break time="350ms"/>
It improves efficiency, <break time="250ms"/>and gets teams live faster.
The second reads like someone making a point. The first reads like a policy document. The difference is timing, not voice.
Three punctuation details worth knowing
- Slashes and pipes are often read aloud or cause odd splits. Replace with a comma or a real
break. - Em dashes and ellipses are unreliable across engines. If you want a pause, write
break. - Parenthetical asides usually wreck phrasing. Promote them to their own sentence or cut them. Listeners won’t hold the aside anyway.
Pronunciation: four strategies, in order of stability
No single method works everywhere. This is the order to try, from most to least stable.
| Strategy | How | Use when | Risk |
|---|---|---|---|
| 1. Rewrite | Swap for an easier synonym | Homographs, rare loanwords | Can shift meaning; needs review |
| 2. Respell | Spell the sound (Sopha → Sofa) |
Brand names, no lexicon available | Display text goes wrong; keep a separate caption string |
3. sub alias |
Keep the word, change the reading | Acronyms, proper nouns, engine supports it | Inconsistent support; audition |
4. phoneme |
Specify phonemes directly | Last resort, single engine | Most fragile; breaks on engine change |
English homographs and how to route around them
| Word | Ambiguity | Safer rewrite |
|---|---|---|
| live | Live stream vs. live here | “real-time”, “broadcast” |
| read | Present vs. past | “reads” / “read earlier” |
| lead | To lead vs. the metal | “leads”, “lead generation” |
| record | Noun stress first, verb stress second | “recording”, “to capture” |
| close | Near vs. to shut | “nearby”, “shut down” |
| wind | Air vs. to twist | “air”, “coil” |
Context usually resolves these. When it doesn’t, add a word — cheaper than a phoneme tag that only works on one engine.
Acronyms and brand names
| Case | Wrong | Fix |
|---|---|---|
| Initialism read as a word | API → “appy” |
<say-as interpret-as="characters">API</say-as> |
| Pronounceable acronym spelled out | SaaS → “S-A-A-S” |
<sub alias="sass">SaaS</sub> |
| Brand stress misplaced | Soundwaver on the wrong syllable |
sub with a hyphenated hint, e.g. Sound-way-ver |
| Version strings | v2.1 → “vee two dot one” |
Rewrite as “version two point one” |
Numbers, dates, units: declare before you read
| Content | Don’t write | Write instead |
|---|---|---|
| Year | 2026 may read “two thousand and twenty-six” |
<say-as interpret-as="date" format="y">2026</say-as>, or spell it |
| Percentage change | -30% → “minus thirty percent” |
“down by thirty percent” |
| Price | $1,299 reads inconsistently |
“one thousand two hundred ninety-nine dollars” |
| Phone numbers | Digit string is unintelligible | Group with pauses: “555, 0100” |
| Time | 15:30 → “fifteen thirty” |
“three thirty in the afternoon” |
Decide the spoken form while writing the script. It’s far cheaper than retrofitting tags across a finished project.
Rate and pitch: small moves only
prosody is the most abused tag. From practice:
- Rate: baseline 100%, work in 90–110%. Below 85% sounds unnatural; above 115% loses clarity. Confirm your baseline first — conversational English lands around 125–165 wpm.
- Pitch: ±5% already separates “statement” from “heads-up”. More than that and it sounds like a different person.
- Volume: don’t mix with
prosody. Handle gain in post.
One trick that reliably works: keep one rate for the whole piece, slow the CTA by 5%. The shift is barely measurable and completely audible.
A six-step workflow that stays maintainable
| Step | Action | Output |
|---|---|---|
| 1 | Write a speakable script; read it aloud before adding markup | Plain script |
| 2 | Add markup only where needed; leave the rest alone | Marked script |
| 3 | Audition three segments on the real engine: opening, proper-noun dense, CTA | Audition log |
| 4 | Keep a mispronunciation list; fix each word once | Brand lexicon |
| 5 | Version per slug, markup and script in sync | slug-v3.ssml |
| 6 | On voice or engine change, re-run the step 3 list | Regression report |
Step 4 is the one people skip, and it decides how much of this you repeat next month. One shared list of problem words covers every future project.
When your tool has no SSML field
Plenty of real workflows give you a plain text box. You can still get most of the benefit:
| Goal | Plain-text substitute |
|---|---|
| Short pause | End the sentence. Stop stringing clauses with commas |
| Long pause | Start a new paragraph, or add a blank line |
| Proper nouns | Rewrite to an easier synonym or a hyphenated spelling |
| Numbers | Write them out as words |
| Rate variation | Generate separate segments, then cut them together |
| Emphasis | Put the key phrase in its own one-line sentence |
The last one — generate in segments, assemble in the edit — is the most useful trick. Opening at 95% rate, product section at 100%, CTA at 90%: three generations, aligned in the timeline. Far more natural than fighting one file.
Pre-publish checklist
- [ ] Every brand, person, and product name is on the fix list
- [ ] No double punctuation used to fake a pause
- [ ] Every number has an explicit spoken form
- [ ] Every acronym confirmed as letters or as a word
- [ ] Opening five seconds, proper-noun section, and CTA auditioned by a human
- [ ] Markup and script share a version number, so a voice change can be re-run
FAQ
Do I need SSML at all? Not always. Short scripts with few proper nouns and modest tone requirements are fine with correct punctuation alone. SSML pays off through repeatability: once the markup is fixed, you can swap voices and languages without rewriting.
Why does the same script mispronounce on a different voice? Voices ship with different lexicons and models. Keep markup attached to the script and auditions attached to the voice. On a voice change, re-run only the proper-noun segments.
IPA or the engine’s native phoneme set?
Check the docs. IPA is broader but partial; some engines only accept their own set. Single-engine projects are safest with the native format. Cross-engine projects should prefer sub.
Can homographs be fixed with phoneme tags? Sometimes, but it’s brittle. A tag applies to the sentence in front of you and silently breaks when the copy changes. Rewriting to an unambiguous word costs less over a year.
Are longer pauses always better? The opposite. Past 700 ms on headphones it sounds like a dropped connection. What works is a short pause in the right place — 300 ms before the key line beats a full second mid-sentence.
How do I do multi-speaker dialogue? Generate each line separately with its assigned voice, then assemble. In-file voice switching is unsupported on most engines and produces inconsistent rate and timing.
Try it
To hear the difference, generate the same script twice at different rates with Soundwaver TTS. If you need one consistent brand voice across every asset, voice cloning keeps the delivery identical. For scripts that are easier to read aloud, see voiceover scripts that convert; for the multi-language side, the localization QA checklist covers what to verify. Plans are on the pricing page.
Note: examples here use standard SSML. Supported tags and behavior differ between engines — audition your key segments on the one you ship with.


