Skip to main content
Ask a question about a recording, then request the audible evidence that supports the answer. Keep what a speaker says separate from conclusions the recording cannot establish. For example, a speaker’s proposed delivery date is not proof that the delivery happened.

Ask about a spoken decision

Install perceptron>=0.4.0, set PERCEPTRON_API_KEY, and put a short WAV recording of a delivery discussion at ./delivery-discussion.wav. Use your own recording and adapt the question to its subject. This example asks for the stated reason for a delay and distinguishes a confirmed date from a proposal:
The response is text. The audio() helper detects WAV, MP3, or FLAC from local file contents and sends the recording inline. See audio input forms for MP3, FLAC, URLs, and uploaded file IDs.

Ask a follow-up with the recording in history

Run this after the first example to ask for the words supporting its answer:
The full messages list sends the original audio and previous answer again. The API does not retain this conversation automatically. Text-only follow-ups do not add media assets; this recording remains asset 0 if a later answer uses an annotation. See asset ordering when combining recordings with other media.

Choose the question and evidence

  • Ask one specific question, such as what decision was stated or which sound occurred before another, and allow an answer of “not audible” or “not stated.”
  • Request a short supporting quotation for speech. For non-speech events, ask for a description of the audible evidence rather than an unsupported explanation of its cause.
  • Use audio clipping when the answer needs an interval to review, or transcription when you need the full spoken text.
  • Use Multilook for independent questions over the same recording. Keep dependent follow-ups in a conversation that includes the earlier answer.
Check the audio and context limits when carrying recordings through several turns. Each request includes the audio supplied in its history.