Repository navigation
Adding emotion/style control when fine-tuning for a new language (Uzbek, custom Latin vocab) #1344
nematov9844
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi, and thanks for F5-TTS.
I am fine-tuning F5-TTS for Uzbek (Latin script) with a custom ~32-character
vocabulary, and I want to add controllable emotion/style on top. Reference
audio alone transfers the speaker's timbre too, so style and speaker get
entangled. I would appreciate guidance on:
from the speaker (e.g. an extra style embedding, or a second reference
used only for prosody)? Is there a recommended place in the DiT to inject
it?
speech in the target language, would the emotion generalize to the target
language, or does the model tie emotion to language/speaker? Any
experience with cross-lingual transfer?
of clean data for natural prosody, fine-tuning from the multilingual
checkpoint vs training from scratch?
duration/ total-length estimation have a recommended approachfor non-English languages? In my tests, a fixed chars-per-second ratio
gives badly timed output for long inputs.
I can share logs or samples if useful. Thanks!
All reactions