xls-r

one script for xls-r/xlsr/whisper fine-tuning

Real-time voice deepfake detection, watermarking & vishing defence (team project, CUD). Server model XLS-R + AASIST (2.84% EER, ASVspoof 2021 LA); separate 0.62 MB ONNX edge model (8.47% EER). Explainable, IPED add-on, FastAPI + React.

Evaluating the Effectiveness of Transformer Layers in Wav2Vec 2.0, XLS-R, and Whisper for Speaker Identification Tasks

Scripts and files I used throughout my Master thesis.

Detect AI voice-clone scams in WhatsApp voice notes — fine-tuned XLS-R 300M, 96% accuracy on unseen voices, AR/FR/EN. GOMYCODE × NVIDIA hackathon 2026.

TFM — Speech Emotion Recognition en español: benchmark LOSO de modelos SSL preentrenados (XLS-R, Whisper, WavLM, HuBERT, TRILLsson, emotion2vec…) sobre EmoMatchSpanishDB. Frozen probing, fine-tuning parcial, fusión y contraste estadístico.

Investigating Layer-Specific Performance in Speaker Recognition with XLS-R Architecture

Fine-tuned XLS-R 300m for Welsh speech recognition — 33% lower WER than zero-shot Whisper-small on FLEURS Welsh

Speaker-independent speech emotion recognition on RAVDESS: MFCC, eGeMAPS and ComParE vs frozen WavLM, HuBERT, XLS-R and Whisper under nested leave-one-speaker-out evaluation, with end-to-end fine-tuning, protocol-inflation and contamination checks, and a per-clip comparison with human listeners.