Minimax Female Accents

作者：imb101

Updated V2 (11/13/26) A single model was sadly mixing all the accents together, so I've split it to a lora per accent. The only downside to this is you can't have two people with different accents talking in one generation. I think the accents sound a lot more accurate and actually have reverb now, though I'm not sure about Utah still. Also added Liverpool and Minnesota. Suggest you use weights between 0.6-0.8 and audio sampling of 3. Prompting is the same as before.

Concept An experimental minimax audio only lora to add some extra variety to specific female accents. While minimax can do some accents they often to tend to sound quite flat, so this just amps them up a bit. Right now I've only had time to add 3 accents; Utah, Atlanta and New Jersey. Hopefully depending on time I can add some more. I will say its far from perfect at getting accents right but it's hopefully an improvement over the base.

Usage Use a weight of 0.4-0.5. if you get odd vocal fry try uping the audio sampling to a higher number, I find 3 is a pretty good sweet spot.

Prompting: Just add 'in a <accent_name> accent to your speech prompt. For example: The woman says <d>[English]Girl, get your fine self over here and give me a hug.. Lord, it's been way too long, come on, sit down and tell me everything</d> in a atlanta accent The woman says <d>[English]Oh, hey.. it's good to see you, I was just thinking about you last night on the drive back from up the mountain.</d> in a utah accentThe woman says <d>[English]Oh my GOD, you're alive!! I been calling you since Tuesday and I was about to call the cops, now get in the car.</d> in a new jersey accent You can give two different people in the same scene different accents but sometimes they can blend together.

Training The model was trained on 66 audio clips, each accent getting 22 clips each, 8 repeats per clip and 20 epochs. This lora doesn't touch the video segments of the model so should have zero impact on the video output. Captioning was automated using Qwen ASR and a basic python script. Training was run using a custom version of diffusion pipe on 4x3090s