Generations out of order, repeating and incomprehensible

[cmdrcc]

3. What OS (Operating System) are you using?

Windows

4. Upload logs.txt

logs.txt

5. Message

bugged speech with human sample.wav
bugged speech no input.wav
bugged out speech gen using previously generated speech sample.wav

Created a voice sample with no problems but then when I used that sample to generate again it was out of order, repeating, and generally making no sense. Ive included logs and the 3 outputs I tried, one with my original audio sample, one with the clone voice and one with no custom voice input at all. All have the same issues.

[francislabountyjr]

@cmdrcc What does the input text look like? I'll run some tests and see if I can replicate

[cmdrcc]

The Man Who Was Definitely Not a Goat

Gerald considered himself a perfectly ordinary man. He woke every morning, adjusted his tie, and trotted off to work at the accounting firm. Everyone said he was good with numbers, though admittedly he did have a strange habit of eating important memos. Gerald always blamed “stress-induced snacking.”

At lunch, Gerald skipped the sandwiches his coworkers ate. He preferred the little patch of grass near the office parking lot. “It’s organic,” he’d say proudly, while they politely pretended not to notice him chewing with his mouth full of dandelions.

On weekends, Gerald joined the local yoga class. The instructor often marveled at how easily he mastered “Downward Dog.” Gerald would correct her: “Actually, I think I’m doing more of an upward… bleat?” Nobody was quite sure what he meant, but he was flexible, so they let it go.

The first real clue came when his barber refused service. “Look, Gerald, I can cut hair, not wool.” Gerald was offended—until he looked in the mirror and saw his reflection: rectangular pupils staring back at him. He blinked. The reflection blinked. He sneezed, and out came a sound no man should make: “Maaaaaaa.”

Gerald was horrified for a moment—then relieved. It explained the grass, the memo-eating, the four-legged urge to climb onto office furniture. “Well,” he said, shrugging as best as a goat can, “at least now I know why I never quite fit in with the guys at the pub.”

And with that, Gerald trotted proudly into the world—finally embracing his true self: a goat in a tie, the best darn accountant in the county.

[francislabountyjr]

@cmdrcc Thank you! Looks like the model has a hard time once the text gets too long. I could generate up to

The Man Who Was Definitely Not a Goat

Gerald considered himself a perfectly ordinary man. He woke every morning, adjusted his tie, and trotted off to work at the accounting firm. Everyone said he was good with numbers, though admittedly he did have a strange habit of eating important memos. Gerald always blamed “stress-induced snacking.”

At lunch, Gerald skipped the sandwiches his coworkers ate. He preferred the little patch of grass near the office parking lot. “It’s organic,” he’d say proudly, while they politely pretended not to notice him chewing with his mouth full of dandelions.

On weekends, Gerald joined the local yoga class. The instructor often marveled at how easily he mastered “Downward Dog.” Gerald would correct her: “Actually, I think I’m doing more of an upward… bleat?” Nobody was quite sure what he meant, but he was flexible, so they let it go.

In one go with good quality. Anything past that caused the issues that you mentioned. I will work on an update for auto-chunking the input text when it is too long. I will make sure the frontend accommodates being able to regen/qc the individual chunks before combining to the final audio.

For now, try to keep generations under ~900 characters/150 words for consistent quality

[devinschumacher]

@francislabountyjr is this solved in our new update im going to release?