RS1_BeatTheta

A compact diffusion model that generates ambient sound textures for theta-range binaural-beat sessions.

RS1_BeatTheta creates the soft background layer (pads, drones, ocean, rain) and can lay an exact binaural beat on top, so a calm, headphone-based listening session takes a few lines of Python.

Model	Class-conditional diffusion, about 5.2M parameters
Output	8 kHz audio: mono texture, or stereo with a binaural beat
Textures	warm_pad · deep_drone · ocean_swell · soft_rain · crystal_pad
Training data	100% synthetic, generated by code
License	MIT

Quick start

pip install transformers torch scipy soundfile

import soundfile as sf
from transformers import AutoModel

model = AutoModel.from_pretrained("Rumiii/RS1_BeatTheta", trust_remote_code=True)   # add .to("cuda") for GPU
audio = model.generate("ocean_swell", seconds=20, beat_hz=6.0)                       # stereo, 6 Hz theta beat
sf.write("session.wav", audio, model.sample_rate)

Leave out beat_hz for the texture alone. trust_remote_code=True is needed because the model code ships in this repo.

Training data

Every training clip is synthesized from code: five texture families, 600 clips each, 3,000 clips of about 2 seconds (roughly 1.7 hours of audio). Each clip is rendered with randomized pitch, detuning, slow modulation, noise colour and filtering, so no two are alike. No recordings or third-party data were used.

Training

Class-conditional diffusion with classifier-free guidance, trained from scratch for 12,000 steps (batch size 24, about 96 passes over the data) with AdamW, EMA weights and mixed precision. About 36 minutes on a single NVIDIA T4.

Intended use and limitations

• Intended for relaxation and focus audio, and as a lightweight base for experiments with synthetic-data audio models.
• Output is 8 kHz; longer audio is built by crossfading short generated clips. Textures only, no music or speech.
• The binaural tones are exact signal processing, not learned. Responses to binaural beats vary between people and the research evidence is mixed. This is a research release (v1), not a medical device or treatment.

Safety

Use stereo headphones at a comfortable volume. Do not listen while driving or operating machinery. Ask a doctor first if you have a seizure disorder or heart condition, or are pregnant.

Hemi-Sync is a registered trademark of The Monroe Institute; this project is independent and not affiliated with it.