Accept, edit, skip, re-record or abandon the question the user is on, and receive the next one.

BCP-47-ish language code (e.g. 'en', 'zh-cn') for localized messages.

The end-user's username.

The end-user's password on first sign-in. After login/register, send the returned session token (JWT) here instead — the field accepts either, and reusing the token keeps raw passwords out of your storage.

Identifies one interview. Returned by start_voice_interview and required by every later call.

Which question this call is about. Send back the one you were last given; it makes the call safe to retry.

What the user did with the question: confirm accepts the transcript unchanged, edit accepts text they corrected, skip moves on without an answer, re_answer discards the recording and re-serves the same question, and abort ends the interview and discards it. skip, re_answer and abort must not carry confirmed_text.

The answer text the user accepted, after any correction they made.

Optional. An id your app makes for this one answer, at most 64 characters, and sends again unchanged with every retry of it. With it, a retry that arrives while the answer is still being saved returns is_processing instead of an error, and one that arrives after it was saved returns the reply the first request was given. Make a new one for each answer.

Whether the call succeeded.

Human-readable reason when success is false.

Identifies one interview. Returned by start_voice_interview and required by every later call.

Which question set to run.

Which question this call is about. Send back the one you were last given; it makes the call safe to retry.

The question the user is being asked, in the interview language.

A time-limited URL for the spoken question. Play it to the user.

Media type of the question audio, so a player can be chosen without sniffing.

How far through the interview the user is, from 0 to 1.

True when there are no further questions and the write-up can be requested.

True when this picked up an interview the user had already started rather than beginning a new one.

True when this request repeated, by its client_answer_id, an answer that is still being saved. Nothing else in the reply describes a question: wait a moment and send the same request again to receive the reply the answer was given. Absent otherwise.

Where the interview itself stands: in_progress while there are questions left, completed once the write-up can be asked for, aborted when the user ended it and the transcript was discarded, and halted_for_safety when an answer suggested the person needed help rather than another question -- on which show safety_message and stop asking. The new first interview instead uses safety_paused: its earlier confirmed answers are kept, the triggering answer is excluded, and questions remain paused.

Present only when the interview stopped because an answer suggested the person needed help rather than another question. Show it, and do not continue the interview.

Present only on an interview made of spoken lines. Which chapter the question is in.

Present only on an interview made of spoken lines. The chapter's title, in the interview language.

Present only on an interview made of spoken lines. Where the chapter comes, counting from 0; a new value means a new chapter has begun.

Present only on an interview made of spoken lines. How many chapters it has.

Present only on an interview made of spoken lines. Roughly how many minutes of answering are left, rounded up, from how long this user's answers have taken so far; 0 once it has ended.

Present only on an interview made of spoken lines. How the user accepts the transcript of their answer to the question: auto lets it through after a short pause unless they touch it, and confirm waits for them, as for a name or a date of birth, and for every answer once they have corrected one. Null when nothing is asked.