lightk/audiodeepfake-detecter-laogao

AudioDeepFake-Detecter (laogao fine-tune)

Binary classifier for synthetic vs. real speech, built on DistilHuBERT. Code: https://github.com/zLiGhTKz/AudioDeepFake-Detecter

Training

1. DistilHuBERT fine-tuned on the ASVspoof dataset for bona fide vs. spoofed speech.
2. Further fine-tuned on recordings of a single speaker ("laogao") and synthetic clips of that voice.

Labels: 0 = fake, 1 = real. Input: 16 kHz mono audio, scored in 1-second windows.

Limitations

• Because of step 2, the model is specialised toward one speaker. It has not been evaluated on other speakers or on public benchmarks, and no accuracy figures are claimed.
• Not suitable as sole evidence in any high-stakes decision.

Usage

from transformers import pipeline
clf = pipeline("audio-classification", model="lightk/audiodeepfake-detecter-laogao")
print(clf("speech.wav"))

Or with the package: audiodeepfake-detect speech.wav --model lightk/audiodeepfake-detecter-laogao.