For audios longer than 1.5-2mins, gpt-4o-mini-tts works really unstable: random pauses from few seconds for more than a minute, random volume and tone changes, repeating last few sentences in random order.
Observing it all the time with very different instructions and inputs. Any known workarounds?
3 Likes
I’ve noticed that too.
Breaking the text into smaller parts of up to 1 minute seems to make it more stable, if that’s a viable option for you.
2 Likes
Same here. Total requested audio was 4:31, but from 1:21-2:26 and 3.02-3:36 there was only silence. Also huge volume level changes and style shifts. In short: unusable crap.
These limitations should be clearly stated.
1 Like
We rolled out a fix today for a class of issues that were causing excessive stretches of silence and repetitions in the output. Could you give it another try and let us know if it’s still happening?
1 Like
It’s still pretty bad. When using 4o mini tts, the voice is coarse. But the older tts was a good smooth voice. Here’s an example of what I mean:
4o-mini-tts (bad) - 4o-mini-tts
older-tts (good) - old-tts-model
I have used the exact same voice onyx. And given no instructions. The only change was the model Id
this issue is still happening in Oct.
I have tried different voices, response formats(mp3, wav, flac etc) , stream formats (sse vs audio)
i have also reduced audio length to 30-45 seconds.
the issue still persists. Any help is appreciated here.
Still unusable in December.
I used ‘alloy’
Had to chunk my script up into two. The first sounded completely like a male. The other sounded like alloy should, but slowly turned darker and darker, at which point artifacts began to appear too.