How to Fix Common AI Voice Problems

Most AI voice problems trace back to one of three things: the sample (echo, noise, flat or fast delivery), the settings (too much emotion or repetition penalty), or the text (spellings, punctuation and line length). Find the symptom below; each comes with its likely cause and the fix that usually works.

The sound of the voice

Symptom	Likely cause	Fix
Echoey, distant, "in a bathroom"	Room echo in the sample	Re-record closer to the mic in a softer room, then replace the clip. See reducing noise.
Hiss, hum or rumble under every take	Background noise in the sample	Remove the noise source and re-record, or clean the sample gently.
Muffled, narrow, "on the phone"	A Bluetooth headset mic, or heavy noise suppression	Record with a wired mic or the phone's own mic.
Doesn't sound like the person	A short, quiet or noisy sample; or you're judging your own voice	Use 5 to 15 seconds of clean speech. Ask someone who knows the voice.
The wrong accent	The voice speaks a different language from its sample	Record a sample in that language, or try guidance strength near 0.

The delivery

Symptom	Likely cause	Fix
Flat, bored, robotic	A flat sample, or Calm with low expressiveness	Record a livelier sample; try Balanced or Expressive.
Over the top, unstable, odd noises	Emotion intensity or expressiveness set high	Go back to Balanced, then raise one slider at a time.
Rushed	A fast sample, or high emotion intensity	Lower guidance strength to about 0.3, add full stops, or record a slower sample.
Every take sounds different	High expressiveness	Lower it to around 0.6 to 0.8. See keeping voices consistent.
Stress on the wrong word	The sentence allows more than one reading	Rephrase so the important word comes last. Capitals don't help.

The words

Symptom	Likely cause	Fix
A word is mispronounced	A name, a rare word, or a word with two readings	Respell it phonetically. See fixing mispronunciations.
Numbers or symbols read oddly	Digits and symbols can be read several ways	Write them as words. See numbers and abbreviations.
Keeps talking past the end, or repeats itself	Repetition penalty set high	Set it back to about 1.2. Runaway pieces are also made again automatically.
A word or phrase is skipped	A long or tangled sentence	Usually caught and re-made automatically. If not, generate again or split the line.
A speaker's name is read aloud	The name before the colon didn't match a voice when pasted	Remove the name from the line's text and pick the voice on the line.
A stage direction is read aloud	Directions or sound cues inside the line	Take them out of the text; set delivery with the style.

The automatic checks

Every take is listened back to with Whisper speech recognition before you hear it. A piece that seems to be missing words, or that runs on far past its text, is made again, up to three tries. That catches most skipped words and runaway endings, but not mispronunciations or delivery you don't like, which is why every take is still worth a listen.

Messages you might see

"The text has changed since this take was made": you edited the line after generating. Generate again before saving, so the audio matches the words.
"This line's voice has changed since this take was made": you gave the line a different voice. Generate again in the new voice.
"The last try didn't work": something went wrong making that take. Press Try again; if it keeps failing, shorten the line.
"The voice server took longer than ten minutes on this line": the line is very long. Split it into shorter lines.
Fast draft won't run: it needs a clip of at least 5 seconds, and English text.

Still stuck?

Try the same line with a different voice. If it's fine there, the problem is the first voice's sample or settings; if it's wrong in every voice, it's the text. That one test narrows almost anything down.

Fixing common AI voice problems: a troubleshooting guide