giangndm/wav2vec2-vietnamese-kmeans-16384

Vietnamese Wav2Vec2 K-Means Codebook (16,384 Centroids)

This repository contains the 16,384 trained FAISS K-Means centroids used to discretize continuous speech representations from the nguyenvulebinh/wav2vec2-base-vietnamese-250h model.

📊 Model Details

• Base Backbone: Wav2Vec2 Base Vietnamese 250h.
• Cluster Count (K): 16,384 centroids (768-dimensional float32 vectors).
• Training Source: Fitted on 2.6 Million speech frames from both audio-pw733-raw and viVoice datasets.

🔗 Linked Dataset

The tokenized feature dataset generated using this codebook is hosted at: 👉 giangndm/audio-confidence-alignment

🚀 How to Load

import numpy as np
import faiss

# Load centroids
centroids = np.load("kmeans_faiss_16384_centroids.npy")
print(f"Centroids shape: {centroids.shape}") # [16384, 768]

# Build FAISS index for fast search
index = faiss.IndexFlatL2(768)
index.add(centroids)

---

Developed as part of the Multi-Project Voice Research Spike under rd-lumi/luvox-vibe-research.