vn-speaker-profiling 0.1.7

Production of Vietnam Post: Vietnamese speaker profiling (gender/dialect) with PhoWhisper/WavLM-family encoders.

Vietnamese Speaker Profiling

Finetune

python finetune.py --config configs/finetune.yaml

Eval:

python eval.py --checkpoint output/best_model --config configs/eval.yaml --test_name clean_test

Infer:

python infer.py --config configs/infer.yaml --audio path/to/audio.wav

Infer (PhoWhisper with HF Space v2 preprocessing)

python infer.py --config configs/infer_pho_hf.yaml --audio path/to/audio.wav

Install from PyPI (inference)

pip install vn-speaker-profiling
vn-speaker-profiling-infer --audio path/to/audio.wav
vn-speaker-profiling-infer --audio_dir path/to/folder --batch_size 8

Publish to PyPI (maintainers)

python -m pip install -U build twine
python -m build
python -m twine upload dist/*

Datasets:

• ViMD Dataset

• ViSpeech Dataset

Pretrained Models:

• WavLM Base+
• Wav2Vec2 Base
• HuBERT Base
• Whisper Base
• PhoWhisper-Base

In Kaggle:

• https://www.kaggle.com/datasets/thanhlamdev/vimd-dataset
• https://www.kaggle.com/datasets/thanhlamdev/vispeech

Architecture:

    Audio -> Encoder (WavLM/HuBERT/Wav2Vec2/Whisper) -> Last Hidden [B,T,H]
                          |
                 Attentive Pooling [B,H]
                          |
                 Layer Normalization
                          |
                     Dropout(0.1)
                          |
          +---------------+---------------+
          |                               |
    Gender Head (2 layers)     Dialect Head (3 layers)
          |                               |
        [B,2]                           [B,3]

Supported encoders:
    - WavLM: microsoft/wavlm-base-plus, microsoft/wavlm-large
    - HuBERT: facebook/hubert-base-ls960, facebook/hubert-large-ls960-ft
    - Wav2Vec2: facebook/wav2vec2-base, facebook/wav2vec2-large-960h
    - Whisper: openai/whisper-base, openai/whisper-small, openai/whisper-medium

Result:

Mô hình	Kích thước tham số	Nhiệm vụ phân loại	Acc. ViSpeech (Clean)	Acc. ViSpeech (Noisy)	Acc. ViMD (Baseline)
wavlm-base-plus	~94 triệu	Gender	96.53%	97.35%	98.66%
		Dialect	88.33%	84.41%	88.49%
wav2vec2-base	~95 triệu	Gender	93.13%	95.59%	98.52%
		Dialect	87.13%	83.63%	88.65%
hubert-base-ls960	~96 triệu	Gender	96.93%	96.67%	98.62%
		Dialect	87.40%	82.55%	87.52%
spkrec-ecapa-voxceleb	~22 triệu	Gender	96.80%	98.43%	N/A
		Dialect	65.33%	65.10%	N/A
PhoWhisper-base	~73 triệu	Gender	95.53%	N/A	98.57%
		Dialect	93.28%	N/A	90.67%
wav2vec2-base-vi-vlsp2020	~95 triệu	Gender	N/A	N/A	98.72%
		Dialect	N/A	N/A	90.61%

Model

https://drive.google.com/drive/folders/1UCOVh9ut8jHmCFfMKgwM2_Mi_3rGhd94?usp=sharing

Citation:

https://github.com/TranNguyenNB/ViSpeech

https://huggingface.co/datasets/nguyendv02/ViMD_Dataset