Audio Conversation Survey Questions

Audio Conversation runs a spoken, AI-moderated interview: the respondent talks, the model asks follow-up questions, and its reply is spoken back. By default it is hands-free — the respondent presses Start conversation once, and from then on the interviewer listens, answers out loud when they stop talking, and listens again, like a phone interview. It suits qualitative probes, guided reflection, or interview-like prompts where a spoken exchange fits the study better than a typed one — the AI-moderated equivalent of Text Chat, but voice in both directions. It requires a Standard plan or above. For an interviewer respondents can see, use Video Conversation.

• System prompt control defines how the assistant should interact with participants, same as Text Chat. It can include piped answers and variables: type {{ to pick from earlier questions and survey variables, and a reference that resolves flashes mint as you close it — the same as in question text.
• An optional spoken greeting lets the assistant open the conversation before the respondent's first turn.
• Three independent pipeline legs — speech-to-text, the reasoning model, and text-to-speech — can each use a different vendor (ElevenLabs, Gemini, Hume, or an OpenAI-compatible endpoint), either directly or through a saved credential. A provider preset fills all three at once; "Custom" leaves each one manual. OpenAI Realtime replaces all three with one live voice model (see below).
• Turn taking is Hands-Free by default: the microphone stays open for the whole conversation and a turn ends when the respondent has been quiet for the End-of-turn pause (1.5 seconds unless you change it; raise it for people who pause mid-thought). There is no button to submit a turn: respondents just stop talking. They can Skip a reply or Pause the conversation, which turns the microphone off. OpenAI Realtime conversations take turns the same way, with no button either. Choose Push to Talk instead if respondents should press to start and stop every turn.
• Let the interviewer end the call (off by default) lets the model end the conversation once it judges the topic is covered, after speaking its closing remark, instead of every respondent running for exactly max turns. Max turns still applies as a hard cap, and min turns is a floor the model cannot end before — the same option Text Chat offers.
• Model choice per step — the transcription model, the reply model, and the voice model — each with suggestions for the selected provider, the same way Text Chat lets you pick its model.
• Turn limits and turn duration (minimum/maximum respondent turns, minimum/maximum seconds per recording) keep conversations comparable across respondents.
• Playback controls choose (for Push to Talk; hands-free always speaks its replies) whether the spoken reply plays automatically and whether a running text transcript is shown alongside the audio. New questions start with the transcript hidden, so respondents just talk; turn on Show transcript if they should see it. The transcript is still stored and exported either way.
• Show when the interviewer is speaking (off by default) adds a line such as "Interviewer is speaking…" while each reply plays. Leaving it off makes the call feel like an ordinary conversation rather than a series of announced machine turns. Interviewer name sets what respondents see the interviewer called in that line and in the transcript labels (blank uses "Interviewer"); it is translated with the rest of the question's respondent-facing text. Both apply to OpenAI Realtime conversations too.
• A typed-reply alternative can be turned on per question — a low-key "type instead" control a respondent can use for any turn, without justification, if they can't provide a spoken reply. The typed-reply box has its own Character limit, Disable pasting, and Track pasting settings.
• Response delay can make the interaction feel paced rather than instantaneous.
• Transcript export includes each turn's text and which model produced each assistant reply, alongside the respondent's own recording. In the Export tab's response table, a conversation has a Play button that plays every recorded turn in order, and each turn's Respondent Audio column plays that turn.

1. Step 1: Open Block Builder. Open the survey and select the block where the voice interview belongs.

   Step 1: Open Block Builder

2. Step 2: Add an Audio Conversation question. Choose Audio Conversation from the question type selector.

   Step 2: Choose Audio Conversation question type

3. Step 3: Write the prompts. Add the respondent-facing prompt, the system prompt that controls the interview, an optional spoken greeting, and any on-screen instructions.

   Step 3: Write the prompts

4. Step 4: Configure the pipeline. Pick a provider preset (or Custom), then choose a saved key for each provider or paste one. A pasted key is saved to Settings → Integrations → API Keys as soon as you leave the field — surveys never store raw keys, so a key that isn't saved can't be used when the survey runs. Set the transcription, reply, and voice models and the voice. For an OpenAI-compatible provider, a blank endpoint uses the address shown in the field, which follows the model (OpenAI for gpt-, whisper-1, and tts- models, Groq for whisper-large and Llama models); only a model from another provider needs its endpoint typed in. Per-step vendors, endpoints, and extra request JSON are under Advanced setup.

   Step 4: Configure the pipeline

5. Step 5: Test the connection. Press Test Connection. Domandata asks the reply model for a sentence, speaks it with your voice settings, then transcribes that audio — so one press checks every key, model, and endpoint. If a step fails you see the provider's own reason (for example a model your key can't use), which respondents never see.

6. Step 6: Set turn taking, turns, duration, and accessibility. Choose Hands-Free or Push to Talk and the end-of-turn pause, decide whether to let the interviewer end the call, then set minimum and maximum turns, minimum and maximum recording length, whether replies auto-play, whether the transcript is shown, whether to show when the interviewer is speaking (and what to call the interviewer), and whether to allow a typed-reply alternative. Instrument Review warns if the minimum turn length is longer than the maximum.

   Step 6: Set turn taking, turns, duration, and accessibility

7. Step 7: Preview with realistic responses. Use Preview to test the interaction before publishing, including brief, off-topic, and hard-to-transcribe (noisy, accented, quiet) recordings.

   Step 7: Preview with realistic responses

8. Step 8: Export and review transcripts. Submit a test response, confirm the transcript and recordings are suitable for coding and audit review, and check privacy expectations before live collection — respondent audio is retained as part of the response.

   Step 8: Export and review transcripts

Choose the OpenAI Realtime provider to run the conversation on a single voice model instead of separate transcription, reply, and speech steps. The model listens continuously and answers out loud as soon as the respondent finishes a thought, and the respondent can interrupt it, so the exchange feels like a phone call rather than taking turns.

• Key: an OpenAI API key — choose a saved one or paste it (a pasted key is saved to Settings → Integrations → API Keys when you leave the field). The key stays on Domandata's server: respondents' browsers connect to the model through it and never receive the key.
• Model and voice: the voice model (default gpt-realtime-2.1) and its voice (default Marin). The Transcripts model writes down what the respondent said, for the stored transcript and exports.
• Replies: Patient waits longer for respondents who pause mid-thought; Quick answers sooner.
• What still applies: the system prompt and opening line, minimum and maximum turns, Let the interviewer end the call (the model ends it by signalling the app, not by saying anything), the transcript display, and the typed-reply alternative. Hands-Free / Push to Talk, the end-of-turn pause, recording length, and auto-play don't apply — the model handles turn taking.
• Pause and resume: Pause hangs up; Resume starts a new call that carries the conversation so far, without repeating the greeting. If the connection drops, the respondent can Resume the same way. Pausing after the last turn ends the conversation.
• Data: each respondent turn is stored with its transcript and a recording from when the interviewer stopped speaking until the respondent finished, and each interviewer reply with its words — the same shape as the other providers, so exports and follow-ups work the same way. A reply the respondent talked over is marked as interrupted, since its stored words may run past what they heard. Replies are capped at a conversational length.

Limits are enforced on Domandata's server. Because the respondent's browser talks to the model directly, Domandata also keeps its own connection to every call (with your key, on the server) and watches it, so a tampered-with browser can't run up your bill:

• The call is hung up once the respondent has used their maximum turns and the last reply has had time to play, or straight away if they try to take another turn.
• Replies the respondent didn't ask for are cut off: a call gets about two replies per turn, and a reply asking for more than the conversational length is cancelled.
• Your system prompt, the reply-length cap, and the end-the-call setting can't be changed from the browser; changes are undone, and a call that keeps trying is hung up.
• Every call has a length ceiling: about two minutes per remaining turn plus one minute, and never more than about 13 minutes. A conversation that runs longer drops, and the respondent can press Resume to carry on with the turns they have left.
• If Domandata can't watch a call, it ends it rather than leave it running. This needs a provider that supports server-side call control — OpenAI's own endpoint does; the endpoint must end in /realtime/calls.

Press Test Connection to check the key, model, voice, and transcript model before collecting responses.

In Settings → Integrations → API Keys, use Add Key under Voice to save a speech provider's key — ElevenLabs, Gemini, Hume, or any OpenAI-compatible service — together with its transcription endpoint and model, speech endpoint and model, and voice. Leave endpoints and models blank to use the provider defaults — for an OpenAI-compatible key, the endpoint that matches the model. When you choose a saved Voice key in the question, its endpoints, models, and voice fill in for the steps that provider serves. A Voice key only signs requests for its own provider; an AI (LLM) key can be chosen for any step.

With a saved key, each step's requests go only to the key's own endpoint or its provider's usual address; an endpoint typed into the step must be on that same server, so nobody who can edit the survey can send your key elsewhere. Changing a saved key's endpoint or provider in Settings → Integrations → API Keys asks for the key again, for the same reason.

A saved AI key's chat endpoint is only used for the reply model. For an OpenAI-compatible provider, Domandata uses the matching /audio/transcriptions and /audio/speech addresses on the same server for the other steps.

Audio Conversation exports the conversation transcript alongside each respondent recording. The assistant's spoken replies are not stored — they're regenerable from the transcript and would otherwise double the size of every response — so only the respondent's own audio is retained. Treat both the recordings and the transcript as qualitative research data, and review privacy, consent, and institutional review requirements before collecting live responses.

• Text Chat Survey Questions for a typed AI-moderated conversation instead of a spoken one
• Audio Response Survey Questions for a single recorded or uploaded clip instead of a back-and-forth conversation
• Set Up Saved Credentials for adding Voice and AI keys
• Content Blocks in Research Surveys
• Create Experimental Conditions
• Publish and Share a Research Survey