S2R (Speech-to-Retrieval) denotes methods that map raw speech input into representations or query signals that can be used to retrieve relevant items from a corpus (text, audio, or multimodal) — either by replacing the conventional ASR (Automatic Speech Recognition) → text-retrieval pipeline or by tightly integrating with it. In practice S2R spans pipeline and end-to-end approaches: the pipeline approach uses ASR output fed to standard IR or dense-retrieval systems, while end-to-end S2R trains speech encoders to produce embeddings aligned with text or document embeddings using contrastive or cross-modal objectives (leveraging ML (Machine Learning) models such as wav2vec 2.0 / HuBERT for speech encoding and DPR/colBERT-style architectures or vector indexes like FAISS for retrieval). For AI practitioners this is significant because direct speech-to-retrieval reduces ASR error propagation for noisy or low-resource conditions, captures paralinguistic information absent from transcripts, enables faster on-device voice search, and allows retrieval-augmented generation workflows to ingest spoken queries or spoken corpora directly; research focuses include cross-lingual and zero-shot alignment, scaling retrieval indexes of speech embeddings, latency/efficiency tradeoffs, and robustness to domain shift.
First used conceptually in spoken-document retrieval research dating back to the 1990s, the specific framing and acronym S2R (and widespread end-to-end neural S2R methods) became popular in the early 2020s as self-supervised speech representations and dense cross-modal retrieval techniques matured and were applied to voice search, multimedia archive search, and retrieval-augmented conversational systems.