#dual-encoder

Google released Speech-to-Retrieval (S2R), a production voice search approach that maps audio directly to embeddings for retrieval and bypasses text transcription, improving retrieval quality across languages.