nyuuzyou's picture
Super-squash branch 'main' using huggingface_hub
ec8119f verified
metadata
language:
  - en
pipeline_tag: text-to-speech
license: apache-2.0
base_model: unsloth/orpheus-3b-0.1-ft
datasets:
  - nyuuzyou/asmr
tags:
  - asmr
  - lora
co2_eq_emissions:
  emissions: 1280
  source: Calculated based on power consumption and regional carbon intensity
  training_type: fine-tuning
  geographical_location: Chelyabinsk, Russia
  hardware_used: 1 RTX 4090 GPU

Orpheus 3B ASMR LoRA

A LoRA adapter for Orpheus 3B trained on ASMR audio data to improve soft-spoken speech generation.

Model Details

  • Base Model: unsloth/orpheus-3b-0.1-ft
  • Training Data: nyuuzyou/asmr dataset (283K clips, 307 hours)
  • Training: 170,000 steps (~40 hours on RTX 4090)
  • Method: LoRA fine-tuning

Capabilities

  • Enhanced soft-spoken speech generation on pre-trained voices (e.g., "tara")
  • Improved gentle vocal characteristics
  • Maintains base model's voice cloning and streaming capabilities

Limitations

  • Not capable of true whispering synthesis - LoRA training insufficient for this complex vocal style
  • Limited ASMR authenticity - cannot generate human-like ASMR content
  • Works best with existing voice profiles rather than novel ASMR characteristics

Ethics

Do not use for impersonation without consent or deceptive purposes.