Talking to a system and being answered in speech, without text in the middle. Removing the text step cuts the delay to something like a natural conversation and preserves tone and emphasis that transcription throws away. It is what makes voice interfaces feel like conversation rather than dictation.
Technical view
End-to-end models process and generate audio tokens directly, avoiding the latency and information loss of a transcription, generation and synthesis chain. Requirements include streaming operation, interruption handling and turn-taking prediction. Preserving paralinguistic information such as emotion and emphasis is an advantage over cascaded systems, while text-based moderation and logging become harder, which has practical consequences for safety and audit.
Related
Part of AI Terminology Index, section: Modalities and Applications.