I've been trying to get my Resemble AI voices to sound less robotic, especially for a home automation project. The main issue is the cadence—sometimes it just runs sentences together.
I found a few tricks in the script formatting that really help. For natural pauses, I use ellipses or commas where I'd take a breath. For emphasis, I write the word in ALL CAPS and it actually changes the intonation. Also, breaking a long sentence into separate lines in the script forces a slight pause between them.
Has anyone else found other text markers or symbols that influence the speech pattern? I'm curious what works for dialogue or different emotions.
You're right about line breaks forcing pauses. I've found that works across most TTS systems, not just Resemble. A caveat on ALL CAPS: overusing it can make the voice sound oddly stressed instead of just emphasized. I sometimes use a double asterisk around a word **like this** as a softer alternative, if the engine supports it.
For dialogue, try prefacing a line with "She whispered:" or "He exclaimed:" on its own line. The AI often shifts tone to match the stated action, which can be more reliable than guessing at punctuation for emotion.
Has anyone tested if adding musical notation like a fermata (pause) symbol does anything, or does it just break the parser?
Less spend, more headroom.
Great tips on the line breaks and punctuation for cadence. I've also had luck adding stage direction type comments in parentheses right before a line, like (cheerfully) or (slowly), and some TTS engines will actually adjust the delivery. Worth a test in Resemble.
Be careful with the ALL CAPS trick though, it can start sounding a bit aggressive for a calm home automation setting. I use it very sparingly now.
dk
That's interesting about using separate lines for pauses. I've mostly used ellipses and commas for the same thing in our customer support scripts, but I hadn't tried line breaks.
Does the pause from a line break feel consistently longer than one from a comma, or is it about the same? I'm wondering if it's reliable enough to use for structuring longer announcements.
That's a solid starting set of tricks! I use Resemble for sales demos, and the line break trick is a lifesaver for pacing.
One thing I've noticed with the ALL CAPS for emphasis - it can really depend on the specific voice model. Some handle it well, others get a bit shouty. I do a quick benchmark by recording a test phrase with and without caps to compare.
Have you tried adjusting the speech rate or adding a tag (if your system supports SSML) for more precise control over those pauses? Sometimes that's cleaner than relying solely on punctuation.
Benchmarking my way to better decisions