aoxo / text2asmr-t3-v2

text2asmr T3 v2 (T2A v1 speech adapter)

A LoRA adapter (r=64, 45M trainable parameters) on the T3 text-to-speech-token model of Chatterbox , trained to deliver ASMR-style speech: slow pacing, whispered and soft-spoken delivery, and explicit pauses and vocal events requested inline in the text.

What it does differently from base Chatterbox

• Inline control tags. Training text carries every pause and vocal event that happens in the audio: [pause 1.5s] , [breathing] , [oral sounds] , [moaning] . The model is never left to invent a pause or a breath; it produces the ones the script asks for.
• Continuation. Each training window is conditioned on the ~4 s of the same recording that preceded it (Chatterbox's speech-prompt slot). Long scripts are generated as chained windows of at most 20 s, each prompted with the tail of the one before, which keeps delivery consistent over minutes.
• Window sizes come from the data. Windows are 2-20 s (the corpus' phrase p99 is 19.8 s), text up to 330 tokens. See docs/GENERATOR_V2.md in the project repository for the derivation.

Training data

• 881,646 windows, 1,931 hours, from ~6,000 creators of long-form spoken ASMR audio ( aoxo/t2a-mommy , aoxo/t2a-daddy ), transcribed with Whisper large-v3 and event-labelled with Qwen3-Omni.
• Creator-capped (no creator above 10 h) and balanced across the two corpora.
• Windows whose transcript loops (Whisper repeating a filler word 4+ times, typically a mis-transcribed moan) were dropped, so the model is not taught to loop.

Evaluation

• Held-out-creator next-token loss (74 creators never seen in training, 20,676 windows): 4.7613 , falling monotonically across the run (4.839 at step 1k).
• Listening evaluation and WER / event-fidelity tests are in progress; this card will be updated with them.

Use

from chatterbox.tts import ChatterboxTTS
from peft import PeftModel
from huggingface_hub import snapshot_download

model = ChatterboxTTS.from_pretrained("cuda")
path = snapshot_download("aoxo/text2asmr-t3-v2", allow_patterns=["adapter/*"]) + "/adapter"
model.t3.tfmr = PeftModel.from_pretrained(model.t3.tfmr, path).merge_and_unload()

For long scripts with chained continuation windows, use scripts/generate_t2a_v1.py from the project repository.

Responsible use

• All output of the reference pipeline is watermarked with Resemble AI's Perth implicit watermark. Keep it on.
• The voice comes from the reference clip you supply. Only use voices you have the right to use, and never to impersonate a real person.
• Training audio is adult-oriented long-form ASMR; the model can produce intimate delivery and vocal sounds.

Limitations

• English only.
• [oral sounds] and [moaning] are rarer in training than [pause] (about 1 event tag per 30 pauses), so their placement is less reliable than pauses.
• Vocal event tags are only as good as the Qwen3-Omni labels behind them; the CLAP v7 classifier that will verify them has not yet met its 90%-per-class release bar.