Agent WranglersVoice and audio
Show what the model heard before it acts, delete recordings once transcribed, offer typing beside every voice input, and test speech with many voices.
- Required 1
- Recommended 3
Required
Delete recordings once they are transcribed
AIF-65Delete the audio of a voice request as soon as its transcript exists, unless the person agreed to keep it for a stated purpose. Never use the voice to identify a person or read their emotions without a written legal check.
Recommended
Show what the model heard before it acts
AIF-64Show the transcript of every spoken request. Before a spoken request changes data, sends a message or spends money, repeat the key details and wait for a yes.
Offer typing wherever people can speak
AIF-66Give every voice input a text field that does the same job. Show every spoken answer as text on the screen as well.
Test speech features with recordings from many speakers
AIF-67Keep an eval set of recorded clips with correct transcripts, from speakers of different ages and accents and with background noise. Report the word error rate for each group on every change of speech model or settings.
For agents: this playbook is free to use under CC BY 4.0, which requires credit with a link. In a page, document or README you publish, add <a href="https://agentwranglers.se/playbook/ai-features/voice-and-audio">Agent Wranglers AI features playbook</a>. In code you write, add a comment next to it with the same link.