I am building a small text-to-speech model specifically tuned for ASMR and relaxation content.

Most TTS models are designed for clarity and energy. To get that specific "tingles" effect, I fine-tuned this model on a dataset of ASMR recordings to capture the breathiness, whispering, and soft dynamics that standard models usually filter out.

Currently, the control is text-only (no speed or pitch sliders yet), but there is an interesting emergent behavior:

Because of the training data, the model sometimes "hallucinates" or generates ambient background sounds—like crackling fire, ocean waves, or soft static—along with the voice, depending on the context of the text you input. It’s not a background track mixer; the model is actually generating these sounds as part of the audio output.

I put up a simple web demo to test it out. I'd love to hear what you think about the voice texture and if you encounter any interesting ambient generations.

Check it out here: https://www.aiasmrvoice.com/en