Type a clean sentence into a text-to-speech engine and it reads back smooth, flat, and a little robotic. Real people do not talk that way. We stall, we backtrack, and we drop in small sounds like um and uh while we think.
Those sounds are called filler words, and they are the fastest way to make synthetic speech sound human. This guide shows you how to add natural filler words to TTS, where to place them, and how to avoid overdoing it.
You will learn what fillers are, why decades of speech research says they help listeners, and the exact steps to write them into a script. The method is simpler than most tutorials claim.
With a conversational model, you skip the special markup. You write the filler where you mean it and let punctuation guide the pauses.
By the end, you can take a robotic line and make it land like a real person on a real call.
What are filler words in text to speech?
You add filler words by writing them straight into your input text, right where a person would naturally pause. Filler words are short words or sounds like um, uh, well, you know, and I mean. People say them while they think or take a turn to speak.
Linguists call these small stumbles disfluencies: the natural breaks, restarts, and sounds that show up in unscripted talk. According to Google Research, disfluency "includes self-corrections, repetitions, and filled pauses (e.g., words like 'umm', and 'you know')."
In text to speech, or TTS, filler words are simply text you add to the script. The synthetic voice then hesitates the way a person does.
This is a deliberate realism lever. It makes a voice agent sound like it is actually thinking.
Why filler words make TTS sound more human
Robotic TTS reads clean prose with perfect pacing and zero hesitation. Real speech is the opposite. We pause, restart, and fill gaps with sound, and listeners expect it.
Fillers are not rare noise. A peer-reviewed survey estimates that natural human-human conversations comprise approximately 5 to 10% disfluencies (Shriberg, 1994). Strip every hesitation out of a script and you remove a signal people rely on to feel like they are talking to someone real.
The data: fillers are the norm, not the exception
The numbers back this up. Linguistics research estimates that approximately 6% of speech is non-pause disfluency (Fox Tree, 1995, as cited). Filled pauses, or filler words, are the most common disfluency in unrehearsed, impromptu speech.
Hesitation also rises with sentence length. That is one reason short spoken sentences read more naturally. They give the voice fewer places to sound stiff.
Fillers help listeners follow and remember
Fillers do more than fill space. In a memory and recall study, filled pauses facilitated recall, while coughs matched in duration impaired it (Fraundorf & Watson, 2011). So the benefit is not just added processing time.
In plain terms, a well-placed uh can help a listener hold on to what comes next. That makes fillers a feature of good speech worth keeping in.
The most common filler words to use
Not every filler does the same job. Some signal thinking, some manage turn-taking, and some soften a reply. Here are the ones worth keeping in your toolkit.
One caution: so and like work as fillers only in some spots. Used well, they sound casual. Overused, they sound sloppy, so place them with intent.
How to add natural filler words to TTS scripts
Here is the part most tutorials skip: how to add natural filler words to TTS in practice. The method is short. You write fillers into the text where a person would hesitate, and you usually do not need special markup.
Rime's prompting guide covers the same principle for its conversational models. The four steps below turn it into a repeatable process.
Step 1: Write the filler into the text, not into a tag
There are two common ways to create a pause in TTS. The first is SSML, or Speech Synthesis Markup Language, where you wrap text in tags like <break time="400ms"/>. The second is to write the filler word itself.
Many modern conversational models, including Rime's Coda and Mist, read the filler straight from the text. You type um where you want the voice to think, and the model handles the rest. No tags required.
Here is the difference in practice:
SSML approach: Sure. <break time="400ms"/> I can help with that.
Plain text: Um, sure, I can help with that.
Browse the Rime voice catalog to pick a conversational voice, then write your fillers directly into the lines you send it.
Step 2: Let punctuation control the pauses
With the filler in place, punctuation sets the timing. Each mark tells the voice how long to wait and how to shape the sound. This is where text normalization matters, because the engine reads your punctuation as pacing before it makes a sound.
PunctuationProsody effectUse it forComma ( , )Short pauseA quick beat between phrasesPeriod ( . )Full stopEnding a thoughtEllipsis ( ... )Hesitant, trailing pauseRare, genuine hesitationQuestion mark ( ? )Rising intonationReal questions
Keep spoken sentences short, under about 25 words. Short lines give the model clean, predictable pacing. Use the ellipsis sparingly, since too many trailing pauses sound uncertain.
Step 3: Place fillers where people actually hesitate
Placement is where naturalness is won or lost. People do not hesitate at random. Rime's linguistics work maps where filler words go, and it turns into four simple rules.
- Start of an utterance: Lead with a filler when a turn begins. "Um, let me check that for you."
- After a repeated small word: Add one after a doubled the, a, or I. "I, uh, I think we can do that."
- Before a long or rare word: Hesitate right before a hard word. "Your total is, um, four hundred dollars."
- Before a word you want to emphasize: A short pause draws the ear. "That is, well, exactly the problem."
Put it together and a flat line comes alive:
Without fillers: Sure, I can help you reset your password. Go to the login page and click Forgot password.
With fillers: Um, sure. I can help you reset your, uh, password. So, go to the login page and click Forgot password.
Step 4: Test how the voice reads it
Never ship a line you have not heard. Synthesize the script, listen, and adjust.
Watch for two problems: fillers landing in odd spots, and pauses that run too long. Move the filler, swap a comma for a period, and listen again. A few passes gets you to natural.
Why placement and restraint matter for voice agents
In a real-time voice agent, fillers do a job beyond sounding human. They manage the conversation.
A quick um signals it is the agent's turn to speak and buys a few planning milliseconds while the answer forms. That timing is part of why fillers matter for conversation latency and smooth turn-taking on live calls.
Don't stack fillers or overdo it
The rule is simple: sprinkle, do not stack. Two ums in a row does not sound human. It sounds like a bug.
Placement carries real weight. In a 2022 Interspeech study, Kirkland et al. (2022) found filler position had the strongest effect on perceived confidence (F=598.35, p<.001).
An utterance with no filler sounded most confident, and one with a medial filler sounded least confident.
Treat that finding as a caution. It was a small study (N=35, synthesized speech), so do not overgeneralize. Still, the takeaway holds: a misplaced filler can make an agent sound unsure.
Pair fillers with other realism cues
Fillers are one lever among several. Natural phrasing, short sentences, and conversational back-channeling like mm-hmm and yeah all add to the effect.
Layer them lightly. A voice that hesitates, acknowledges, and keeps its sentences short feels like a real person on the line.
Frequently asked questions
What are filler words in text to speech?
Filler words are short sounds like um, uh, and well that you write into a TTS script so the voice hesitates like a person.
Do I need SSML to add filler words?
Not with modern conversational models like Rime's Coda and Mist. Write the filler directly into the text and let punctuation handle the pauses.
What are the most common filler words?
The most common ones are um, uh, well, so, you know, I mean, and like.
Can you use too many filler words?
Yes. Stacking fillers sounds broken, and a misplaced filler can make a voice sound unsure, so sprinkle them lightly.
Start adding natural filler words to TTS
Fillers are a normal part of speech. They show up in most real conversations, and they make synthetic speech feel human.
The fastest way to add natural filler words to TTS is the simplest one. Write the filler into the text where a person would hesitate, let punctuation set the pace, and place each one with intent. Skip the SSML tags and keep your sentences short.
Try it in a conversational model, listen to how it lands, and adjust. That loop is how a robotic line becomes a voice people trust.
More from the lab
Guide: Design and run an A/B test of your AI voices on real calls using Claude and Cursor
Anything can sound great in a demo, but what do your callers actually prefer? Test multiple voices and vendors in your own experiment with this guide. Get benchmarks to measure against and prompts that you can execute in Claude and Cursor today.
