joeygambino/ltx23-accent-american-audio-lora

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.