Pinyin Semantic Recognition via Vector Embedding

The patent introduces Pinyin sequences as an intermediary representation between speech signals and text information. Instead of directly converting speech to text, the system first converts speech to Pinyin sequences, then uses word embedding models to map these sequences to vector spaces where semantically similar words are distinguished. This intermediary step resolves the ambiguity of homophones while maintaining system feasibility.

The patent transforms the representation parameters of speech recognition from direct text mapping to Pinyin-based vector space mapping. By changing the parameter space from discrete text labels to continuous vector representations with semantic differentiation, the system can distinguish between words with similar pronunciations but different meanings, thereby improving recognition accuracy.

The patent implements a self-service mechanism where the word embedding model automatically learns and distinguishes semantic differences between Pinyin sequences without requiring manual annotation. The model trains on unlabeled data and autonomously develops the capability to differentiate homophones based on contextual patterns, eliminating the need for expensive and time-consuming manual semantic labeling.

The patent uses word embedding to create vector space copies of Pinyin sequences, where each Pinyin sequence is represented as a vector that captures its semantic meaning. This copying approach allows the system to work with semantic representations rather than raw text, enabling automatic differentiation of homophones through vector similarity calculations without manual intervention.