LTX-2.3 American Accent LoRA — audio-branch, rank 32
Makes accent wording in your prompt actually work. LTX-2.3's voice prior
ignores accent requests in exactly the regions where you need them most —
young female characters default to Australian-leaning voices even when the
prompt says "in a casual American accent". This LoRA turns that wording into
a reliable control.
Read this first: the 24 fps rule
This LoRA cannot help you at the wrong frame rate. LTX-2.3's joint
audio-video prior is 24 fps-native, and render fps is a hidden accent dial:
at 25 fps the same prompt and seed render non-rhotic southern British, at
30 fps broad Australian — and off-24 fps overrides accent wording
entirely, LoRA or no LoRA. Set your workflow's fps/frame_rate to 24,
then use this LoRA. (Dose-response verified by A/B on identical configs,
2026-07-30.)
What it does — measured
A/B matrix on the hardest region (young pale-skinned woman, casual camcorder
monologue), 24 fps, two seeds per cell, blind phonetic review:
| spoken-line wording | without LoRA | with LoRA @ 1.0 |
|---|
| no accent stated | Australian | Australian |
| "saying in a casual American accent" | Australian (both seeds) | General American (both seeds) |
| rich scaffold ("her voice is a low American voice with a slight rasp, speaking in a casual American accent") | Australian (both seeds) | General American |
The pattern: the LoRA does not force American unconditionally — it makes the
model obey the accent you ask for in regions where the base model
refuses. State the accent on every spoken line ("in a casual American
accent"); the LoRA does the enforcing.
Lip-sync safe by construction
1,152 LoRA tensors, all of them in the audio branches — audio_attn1 (480),
audio_attn2 (480), audio_ff (192). Zero video tensors, zero cross-modal
(video_to_audio / audio_to_video) tensors. Video content, faces, and lip
motion are mathematically untouched.
Usage
- Strength 1.0 in any LTX-2.3 LoRA loader. Verified in long multishot
production runs alongside a video-branch LoRA with no interference.
- Compatible with the LTX-2.3 22B family: stock distilled 1.1 and the
JoyAI-Echo surgical merges (bf16 / fp8 / INT8 / GGUF via fuse-at-load).
- Render at 24 fps. Say the accent. That is the whole recipe.
Training
- ai-toolkit, rank 32 / alpha 32, 3,000 steps, lr 1e-4, batch 1, qfloat8.
- Data: American-English read speech from LibriSpeech (CC BY 4.0;
V. Panayotov, G. Chen, D. Povey, S. Khudanpur), muxed over static video so
only the audio lane carries signal. Captions are verbatim transcripts with
the accent deliberately unnamed — the American prior trains as always-on
behavior, not a trigger phrase.
- The audio-branch-only module scope is what keeps static training video
harmless: the video branches never receive a gradient.
License
LoRA weights follow the LTX-2 Community License of the base model. Training
audio: LibriSpeech, CC BY 4.0.
