Continued utterances
Speech-to-text ends an utterance at every pause, so a user who pauses in the middle of a sentence – "Give me an appointment" … "with a plastic surgeon" – reaches the agent as two utterances. By default the agent answers each one in turn: it plays its response to the first part, perhaps "Do you mean our booking service?", and only then responds to the second.
With continued utterances enabled, when the user continues speaking before the agent started playing its response, that response is withheld – it is never played – and the agent answers both parts as one utterance: "Give me an appointment with a plastic surgeon". The withheld response is removed from the conversation history the Large Language Model sees, so it does not assume the user heard it.
Enable it by adding the continued_utterance parameter to the Advanced tab of the Agent's or Flow's dialog:
{
"continued_utterance": {
"mode": "enabled"
}
}
This applies to voice calls with Agents and Flows on regular (non-speech-to-speech) models, except in Agent assist mode.
- Only a response the user has not started hearing is withheld. Once any of it played, the next utterance is answered on its own – use Interrupted responses to tell the model how much of it the user heard.
- The agent does not stop the work behind the response: the Large Language Model finishes it, and any tool it called still runs. A tool call and its result stay in the conversation history, so the model knows the tool ran; only the words the user never heard are removed.
- A response that ends or transfers the call, or plays an audio file, is never withheld.
- When the withheld response answered an event rather than speech – for example, a no user input event – the user's words are answered on their own.
- In a Flow, only a response that keeps the conversation on its node is withheld. A node's first response always plays, and so does a response after which the conversation moves to another node – the user's words that came in the meantime are then handled as they would be without continued utterances.
Each withheld response and each merged utterance is recorded as a continued_utterance entry in the conversation log, so the log still shows what the agent would have said.
| Parameter | Type | Description |
|---|---|---|
continued_utterance
|
ContinuedUtterance | Answer the whole utterance when the user continues speaking before hearing the response. |
ContinuedUtterance
| Parameter | Type | Description |
|---|---|---|
mode
|
enum | Supported values: • disabled – each utterance is answered in turn (default)• enabled – a response the user has not started hearing is withheld when they continue speaking, and the whole utterance is answered |
In a multi-agent topology, a sub-agent – an Agent or a Flow – without its own continued_utterance follows the root agent's.