Glossary
Voice AI, in plain language
A shared vocabulary helps architecture, product, operations and risk teams make the same decision for the same reason.
Automatic speech recognition
ASR converts spoken audio into text. Accuracy depends on audio quality, accents, domain language, latency and how the transcript will be used.
Text to speech
TTS generates spoken audio from text. Voice quality matters, but so do latency, pronunciation control, interruption handling and regional availability.
Voice activity detection
VAD estimates when a person starts and stops speaking. It influences responsiveness, interruption and the feeling of a natural conversation.
Barge-in
Barge-in lets a caller interrupt generated speech. Reliable interruption requires coordination between audio detection, orchestration and the conversation state.
Turn taking
Turn taking is the control of when the caller and system speak. Poor turn handling creates overlap, silence and accidental interruption.
Grounding
Grounding constrains an AI response to trusted information or tools. It reduces unsupported answers but does not remove the need for evaluation and controls.
Handoff
Handoff moves the interaction to a person or another service with the right context, reason and state. A transfer without usable context is not a successful handoff.
Evaluation
Evaluation measures task success, accuracy, safety, latency, customer effort and downstream rework across realistic conversations.
Apply the vocabulary
Use these terms in architecture decisions, test plans and supplier discussions.