The LoRA model has been released for generating ASMR audio based on Minimax H3

Developers presented asmr-trigger-audio-h3-lora — a specialized adaptation for the Minimax H3 model. The tool is designed to generate high-quality audio tracks with whisper effects and specific ASMR triggers.

To achieve optimal generation quality, the authors recommend using weights in the range of 0.4 to 0.8. The number of inference steps should be between 15 and 30. The model has undergone initial testing by the community, receiving a rating of 11 stars on the Hugging Face platform within the first hour of publication.

A demonstration video is available in the project repository, showcasing the model's operation with the given parameters. The use of LoRA allows adapting the basic H3 architecture to narrow audio tasks without the need for full retraining of weights, which reduces the requirements for computational resources when generating specific content.

This release expands the possibilities of applying generative models in the field of creating sound effects and audio content. The integration of LoRA adapters into workflows allows users to precisely adjust the stylistics and sound character of neural networks, while maintaining the model's basic speech synthesis capabilities.