Voice AI Glossary: The Terms That Matter When Buying an Assistant

Voice AI products use overlapping terms for very different things. This glossary connects each term to a decision or a piece of evidence you can ask for. It is designed for someone buying or setting up an assistant, rather than compiling search phrases.

Speech and conversation

Speech-to-text, or STT: converting audio into written words. A useful evaluation checks the business details the assistant must capture, such as names and corrected dates, as well as the transcript. See the recognition comparison.

Text-to-speech, or TTS: converting text into spoken audio. Voice quality, pronunciation and timing depend on the selected model and voice. A clear sample is useful, but a conversation also needs interruption and playback handling. Use the TTS worksheet.

Speech-to-speech: a system that can take audio and produce audio without requiring your application to assemble separate transcription and synthesis stages for every turn. It still needs business instructions, permissions and action execution. OpenAI describes this architecture in its Realtime guide.

Turn detection: deciding when a caller has finished enough of an utterance for the assistant to respond. A pause may mean the caller is done, thinking, or giving a number in pieces. Test realistic pauses rather than equating silence with completion.

Interruption or barge-in: a caller speaking while the assistant is talking. Evaluate whether outdated audio stops, whether the new request is understood and whether any already-started action needs to be cancelled or reconciled.

Latency: a delay with a defined start and end. “Time to first synthesized audio” and “time from the caller finishing to the next audible answer” measure different portions of a conversation. Ask for the timing definition before comparing figures.

Instructions and actions

Prompt: instructions and context supplied to a model. Approved facts and clear boundaries help, but a prompt does not create credentials or prove that an action succeeded.

Workflow or graph: explicit conversation stages and transitions. It can make a multi-step task easier to inspect. A useful stage has an objective and completion condition and can reuse earlier information.

Tool or function calling: the model requests an application action with structured inputs. Your application must validate and execute that request and return the actual result. “The model called a tool” is not the same as “the customer was successfully booked.”

Delegation: handing part of the work to another model or application component. For example, GPT-Live can handle voice interaction while a backend handles task reasoning. The application retains responsibility for business records and permissions. OpenAI's migration guidance

Idempotency: making repeated delivery of the same operation safe from duplicate effects. Ask what happens if a booking request times out and is retried, including how the application distinguishes a new caller request from a repeated submission.

Telephone connectivity

PSTN: the public telephone network. A browser conversation does not exercise the same route as a caller using a mobile or landline number. Verify the actual intended path before relying on it.

SIP trunk: a connection used to carry telephone calls between a carrier and a voice system. Number ownership, authentication, routing and outbound permissions still need configuration. See the BYO SIP introduction.

Cold and warm transfer: a cold transfer connects the caller onward without a prior consultation; a warm transfer includes a consultation before the handoff completes. Ask what happens when the destination declines or never answers, and which call legs remain billable.

Testing and money

Simulation: a test where some participants, inputs or external actions are modeled instead of performed. It is useful for checking logic. It does not prove audio quality, telephone reachability or an actual calendar write.

Reservation: a temporary allocation of funds or capacity before an operation. Settlement records the usage afterward and reconciles the hold. Read the final charge rather than treating every reservation as spent money.

BYO credentials: using your own provider account or keys. This can move some charges to that provider; it does not imply the platform, carrier or every other component becomes free.

Successful outcome: the business result you defined and can verify. A completed call, an automated evaluation score and a customer's feedback are different signals. Keep them separate in your reports.

Put these terms into practice with the project planning guide and demo checklist. For Burki-specific charges and allowance, use pricing and the trial guide.