Now live! Update from Pinokio (Update in the menu), then click Start!
⭐ NEW: Quality Mode
Tick Check every word before playing (Quality Mode), now in its own box between Voice and Sampling. Whisper checks every take, and each chunk of a long script, against your script. If words are missing or a direction was read aloud, the take is remade (up to twice) and the best one plays. It plays once it's made instead of streaming.
Under it, Voice guidance (0–3, default 2) pulls a cloned voice closer to its clip. On our test clips, speaker similarity rose from 0.45 to 0.58 and from 0.62 to 0.68, pitch moved closer to the original, and every word was kept. It costs about 1.5× the time per take.
OTHER FEATURES
The Need for Speed: The most-used experts now stay in VRAM, so a 5090 runs at 2.3–2.75× real time (was ~2×), and 24 GB cards get faster too.
Dolly the Sheep: Cloned voices now actually whisper inside <whisper>…</whisper>, and <shouting> or <crying> really change the delivery. Whispered spans went from 2 to 9 out of 10 takes.
Fact Check: Takes are checked while they stream: a skipped opening restarts before you hear it, and missing last words are added at the end. A directed line went from 3/10 to 10/10 complete takes.
Simon & Garfunkel now speak: Takes no longer get stuck in pauses broken by quiet noises (7 of 8 stalled before, 0 of 8 now).
I don't know what to say here, I am just going to say Brazil: In long scripts, a cue at the end of a chunk stays with the words after it, and your own "Direction:" carries across every chunk.
Back to the Future of the Past: Click any earlier take to select it without playing it, then retake it, edit it, or keep its voice.
Smaller quality of life fixes.
Windows and Mac Install fixes.
---
Examples of the new quality mode:
Voice Cloning of Christian Bale, Dark Knight for @morpheus :
I also made this funny parody of the opening of the TV Show "Nathan For You" with Donald Trump instead (the background song I added in post):
Now live! Update from Pinokio (Update in the menu), then click Start!
⭐ NEW: Quality Mode
Tick Check every word before playing (Quality Mode), now in its own box between Voice and Sampling. Whisper checks every take, and each chunk of a long script, against your script. If words are missing or a direction was read aloud, the take is remade (up to twice) and the best one plays. It plays once it's made instead of streaming.
Under it, Voice guidance (0–3, default 2) pulls a cloned voice closer to its clip. On our test clips, speaker similarity rose from 0.45 to 0.58 and from 0.62 to 0.68, pitch moved closer to the original, and every word was kept. It costs about 1.5× the time per take.
OTHER FEATURES
---
Examples of the new quality mode:
Voice Cloning of Christian Bale, Dark Knight for @morpheus :
I also made this funny parody of the opening of the TV Show "Nathan For You" with Donald Trump instead (the background song I added in post):