> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.jambonz.org/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.jambonz.org/_mcp/server. # Conversation The **Conversation** section sets how a call runs once it connects. Some settings apply to every agent; the turn-taking settings apply only where jambonz runs the turns. | Setting | Cascaded voice pipeline | Realtime voice API | Connected agent | Custom harness | | -------------------------------------- | ----------------------- | -------------------------- | --------------------------------- | -------------- | | Answering machine detection | yes | yes | yes | yes | | Noise isolation | yes | no | no | yes | | Background sound | yes | yes | yes | yes | | Recording and session data | yes | yes | yes | yes | | Call goal | yes | yes | yes, except **The model decides** | yes | | Turn-taking, interruptions and silence | yes | no: the vendor takes turns | no: the platform takes turns | yes | | Tool filler | yes | no | no | no | | Call settings (advanced) | yes | no | no | yes | The greeting, and whether the agent waits for the other party to speak first, are set on the [Prompt](/agent-studio/building/prompt#greeting) section. ## Answering the call New outbound agents start with answering machine detection on, Krisp noise isolation and a call goal. New inbound agents start with all three off. Change any of them at any time. ### Answering machine detection **Answering machine detection** tells a person from voicemail before the agent speaks. It is also useful when a dialer places the call and sends it to jambonz as inbound. Its fold sets: | Setting | Options | Default | | -------------------------- | ------------------------------------------------------------------------------------------------------------- | ------------------------------ | | **When a machine answers** | **Leave a message after the beep** (with the **Message to leave**), **Hang up**, **Connect the agent anyway** | Leave a message after the beep | | **When a person answers** | **Connect the agent**, **Transfer to a person** | Connect the agent | | **Give up deciding after** | 3 to 20 seconds | 8 seconds | **Transfer to a person** uses the destination set in [Transfer & hangup](/agent-studio/building/transfer-and-hangup). It cannot dial a destination chosen by the agent, because there is no agent to choose one yet. ### Noise isolation **Noise isolation** cleans up the caller's audio before it reaches speech recognition, which helps when people answer in noisy places. Choose **Off**, **Krisp**, **RNNoise (open source)** or **ai|coustics**. ## Background sound **Background sound** plays a quiet audio loop under the call, such as office ambience, from your own file. Give a public `https://` **Audio URL**; nothing plays until it is set. The fold sets **Loop the sound** (on by default) and the **Volume**, from -60 dB to +6 dB relative to the file's own level (-12 dB by default). ## Recording and session data **Recording and session data** keeps the audio and the full session file for this agent's calls, for playback, downloadable transcripts and call analysis. The switch controls recording; whether session files are kept follows the account's observability setting. The switch needs observability set up on the jambonz account. When it is off, calls are still listed and transcribed live, but nothing is kept. Members cannot change it. ## Call goal **Call goal** says what counts as this agent succeeding. Every call is then marked reached or not reached, alongside who answered (person, machine, no answer). The goal is reached when any of the rules you tick happens: * **A successful transfer to a person**: the person answered and the call was bridged. * **One of these tools ran successfully**: pick the tools that mean the job was done. * **The model decides**: gives the model a `mark_goal_reached` tool. Describe when to call it; the description is added to the prompt. ## Turn-taking For Cascaded voice pipeline and Custom harness agents, a sentence at the top of the section sums up the current settings: how much speech interrupts the agent, when it checks in after silence, and when it gives up. Click a value in it to go to that setting. | Setting | What it does | Default | | ---------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------- | | **Speech needed to interrupt the agent** | how long the caller must speak before the agent stops talking, 0.2 to 2 seconds. Shorter feels responsive; longer ignores coughs and noise | 0.5 s | | **End-of-turn detection** | who decides the caller has finished: the speech-to-text vendor, or the **Krisp turn model** | the speech-to-text vendor | | **Check in after silence** | after this many seconds of silence the agent says something, 0 to 30. Off at 0 | 10 s | Deepgram Flux, AssemblyAI and Speechmatics detect the end of a turn themselves. With another speech vendor and no Krisp, Studio notes that replies will be slower. With Deepgram Flux, three more sliders appear: **End-of-turn threshold**, **Eager end-of-turn threshold** and **Silence that ends the turn anyway**. Left unset, Deepgram's own defaults apply. The fold below holds the rest: | Setting | What it does | Default | | ------------------------------------- | --------------------------------------------------------------------------------- | -------------- | | **Unanswered prompts before hang-up** | after this many check-ins with no answer, the agent ends the call. 0 means never | 2 | | **Let callers interrupt the agent** | the agent stops talking when the caller speaks | on | | **Stay quiet after an interruption** | when off, the agent may resume if the caller only made a noise | off | | **How interruptions are detected** | **Voice activity**, or **Interrupt prediction** with a **Prediction sensitivity** | Voice activity | | **Respond** | with the Krisp turn model: how soon the agent answers, from sooner to later | 0.5 | A Custom harness decides for itself what to do after silence: jambonz tells it when the caller has been quiet for **Check in after silence**, so **Unanswered prompts before hang-up** is not shown for it. **Restore all defaults** at the bottom of the section resets the turn-taking, interruption and tool-filler settings. It leaves answering, background sound, recording, the call goal and the advanced settings as they are. ## Tool filler **While tools run**, in the same fold, fills the silence while a tool is working, on Cascaded voice pipeline agents: * **Off** * **Short phrases** (the default): the agent says brief phrases in the **Phrase style** you describe ("friendly and brief" by default) * **Audio loop**: plays the audio at a public `https://` **Audio URL** **Start filler after** sets how long a tool must run before the filler starts, 0 to 5 seconds (2 by default). An audio loop with no valid URL blocks publishing. Each HTTP tool can also have a line the agent says before calling it. See [Tools](/agent-studio/building/tools). ## Call settings (advanced) **Call settings (advanced)** takes any other call setting jambonz accepts, as a JSON object of up to 8 KB, for example `{"boostAudioSignal": 3}`. Keys jambonz does not accept are listed under the box and left out. Answering machine detection, recording and noise isolation are set by the cards above, so those keys are left out here too.