Spoken QA

Spoken question answering studies how systems answer questions delivered as speech while preserving the meaning, language, and evidence in the input. Approaches range from automatic speech recognition followed by language models and retrieval to end-to-end speech-language models, using speech-text alignment, audio representations, attention-based grounding, and synthetic speech for training. Current work also uses task-specific and on-policy distillation to balance recognition, text reasoning, and speech output, and evaluates systems on factoid, multi-hop, multilingual, noisy-audio, and domain-specific benchmarks. A central tension is whether direct audio processing can avoid transcription errors without sacrificing text-native reasoning, while benchmark coverage and automated judging remain unreliable across languages and answer formulations.