Residual Vector Quantization for Audio and Speech Embeddings

3-Second Blueprint (English)

1. Maps continuous encoder audio frames onto learned discrete codebook indices, storing each vector using log2(N) bits per entry.
2. Iteratively quantizes the residual error vector across cascaded codebooks to reconstruct high-dimensional speech at 1.5–6 kbps.
3. Prevents dead codebook entries during neural codec training by reinitializing unused centroids onto active encoder embeddings.