Bridging the Gap Between Continuous and Informative Discrete Representations by Random Product Quantization

Self-supervised learning (SSL) has become a core technique in speech processing, but the high dimensionality of its representations makes discretization essential for improving efficiency. However, existing discretization methods still suffer from significant information loss, resulting in a notable performance gap compared to continuous representations. To overcome these limitations, we propose two quantization-based discretization methods: Product Quantization (PQ) and Random Product Quantization (RPQ). PQ partitions the original feature space into multiple subspaces and independently quantizes each sub-vector, producing a fused set of discrete units that retain diverse information from different subspaces, thereby mitigating the loss associated with single-cluster quantization. RPQ further enhances representation diversity by randomly sampling feature dimensions multiple times to construct sub-vectors, thereby better capturing the variability in the data distribution. Theoretical analysis shows that RPQ reduces the correlation$\rho \ (0 \le \rho \le 1)$between sub-quantizers, and its quantization error is lower-bounded by$\rho \varepsilon _{\text{kms}}$, where$\varepsilon _{\text{kms}}$is the error of a single K-means quantizer. Experimental results show that, on the combined dataset constructed from LibriSpeech and ML-SUPERB, PQ and RPQ outperform standard K-means discretization, achieving relative improvements of 21.8% and 20.0% in word error rate (WER) on LibriSpeech, and 24.1% and 19.6% in character error rate (CER) on the ML-SUPERB, respectively. Moreover, their performance is competitive with, and in some cases even surpasses, that of continuous SSL representations.