Many bug fixes and smaller issues fixed! Also I added two new opt-in options: Experimental Changes and Length Hint! Please check out if you want, might really improve your generation!
⚗️ Enable Experimental Changes (opt-in)
One new checkbox switches on sampler fixes that others have found since Drama 3's release:
- Uncut main token: the main token is drawn with no top-p cut, so a drone can end. Top-p 0.9 was cutting off the end-of-speech token on every step (torenyx/drama3-24gb).
- Looser stall retry: a take that stalls is remade with a looser repeat-avoiding draw (same source).
- SGLang Omni's fixes of 2026-10-20: the acoustic codebooks are capped at 30 candidates, and the redraw keeps the token that crosses top-p.
- Register guard: in Quality Mode, a take where one phrase jumps far above the voice's pitch is made once more (from drama-tensorfold).
The first two changes (the uncut main token and the looser stall retry) were tested together on pause-heavy lines, with 80 takes a side and two cloned voices:
- Drones: 70 instead of 136.
- Right first time: 61 takes instead of 51.
🔀 Crossfade chunk joins (on by default)
Long scripts no longer play a burst of hiss between chunks. Each chunk keeps full level just past its last word, then only its trailing breath and hiss fade out under the next chunk's first words. No word is ever faded in or out. In blind tests it beat the old trimmed end with a silent pause at 5 of 6 joins.
📏 Length Hint (opt-in)
Tells Drama 3 roughly how long each line should take, using its own [audio tokens: N] hint. That gives steadier pacing, and in blind tests it was preferred on long takes.
✍️ Missing words made
When a take or a long script's chunk stops early or stalls, its missing words are made as a continuation of it. That now works while streaming too, and in Quality Mode it's tried before remaking the whole take.
Bug fixes
- Paragraph breaks fixed: blank lines in a script no longer make the model go silent instead of ending.
- Babble ends sooner: runaway takes are cut at 2.5× their expected length instead of 3×.
- Cleaner chunk ends: with crossfading off, chunk ends keep their last consonant (an 80 ms hold before the fade).
Thanks to SGLang Omni, Frosty40, torenyx, plotarmordev, gabacool and mlx-speech, whose findings these are built on.
Many bug fixes and smaller issues fixed! Also I added two new opt-in options: Experimental Changes and Length Hint! Please check out if you want, might really improve your generation!
⚗️ Enable Experimental Changes (opt-in)
One new checkbox switches on sampler fixes that others have found since Drama 3's release:
- Uncut main token: the main token is drawn with no top-p cut, so a drone can end. Top-p 0.9 was cutting off the end-of-speech token on every step (torenyx/drama3-24gb).
- Looser stall retry: a take that stalls is remade with a looser repeat-avoiding draw (same source).
- SGLang Omni's fixes of 2026-10-20: the acoustic codebooks are capped at 30 candidates, and the redraw keeps the token that crosses top-p.
- Register guard: in Quality Mode, a take where one phrase jumps far above the voice's pitch is made once more (from drama-tensorfold).
The first two changes (the uncut main token and the looser stall retry) were tested together on pause-heavy lines, with 80 takes a side and two cloned voices:
- Drones: 70 instead of 136.
- Right first time: 61 takes instead of 51.
🔀 Crossfade chunk joins (on by default)
Long scripts no longer play a burst of hiss between chunks. Each chunk keeps full level just past its last word, then only its trailing breath and hiss fade out under the next chunk's first words. No word is ever faded in or out. In blind tests it beat the old trimmed end with a silent pause at 5 of 6 joins.
📏 Length Hint (opt-in)
Tells Drama 3 roughly how long each line should take, using its own [audio tokens: N] hint. That gives steadier pacing, and in blind tests it was preferred on long takes.
✍️ Missing words made
When a take or a long script's chunk stops early or stalls, its missing words are made as a continuation of it. That now works while streaming too, and in Quality Mode it's tried before remaking the whole take.
Bug fixes
- Paragraph breaks fixed: blank lines in a script no longer make the model go silent instead of ending.
- Babble ends sooner: runaway takes are cut at 2.5× their expected length instead of 3×.
- Cleaner chunk ends: with crossfading off, chunk ends keep their last consonant (an 80 ms hold before the fade).
Thanks to SGLang Omni, Frosty40, torenyx, plotarmordev, gabacool and mlx-speech, whose findings these are built on.