Task Testing
Part of AI meeting tools
Testing speaker attribution with a sample recording
Use a known-speaker recording to check wrong, missing, split and merged labels before relying on meeting notes.
Test speaker attribution using an authorised recording with speakers and turns you already know. Check whether a tool splits one person across several labels, merges different people, or assigns someone’s words to the wrong person. Wrong attribution matters most where notes assign a decision or task.
Prepare a known-speaker sample
Use volunteers and an invented, non-sensitive discussion for an initial exercise. Obtain the permissions required to make and upload the recording. Make an answer key recording the correct person for every turn, including short replies, interruptions and speech from a shared room microphone. Keep the key outside the tool.
Choose a capture setup resembling the one you intend to use. Separate participant audio streams may give a tool different clues from one room microphone. Include ordinary turns, overlap and a speaker change after a pause. Record the microphone arrangement, spoken language, meeting display names and any voice-profile settings.
Separate a label from a verified identity
A label such as “Speaker 2” indicates a distinct speaker label; it does not identify the person. Save the uncorrected output before editing it, so the test measures what the tool produced.
In Microsoft Teams, a participant can hide their name in captions and transcripts. In a Teams Room without speaker recognition, Microsoft says in-room audio may be attributed to the room in AI notes.
Treat a room or anonymous label as unidentified for a named-person test. The result depends on the room equipment, recognition settings and participant choices.
Review each turn
Compare the output with the recording and answer key. Use these verdicts:
| Verdict | Meaning |
|---|---|
| Correct person | The turn is assigned to the right known speaker. |
| Wrong person | The words are assigned to someone else. |
| Unidentified | The words are present, but the person is unknown, numbered or attributed only to a room. |
| Split or merged | One person receives several identities, or several people share one identity. |
| Missing or unusable | The turn cannot be checked from the output. |
Report affected turns alongside the total turns reviewed, and identify errors involving decisions, objections or accepted actions. Record the correction work before sharing the transcript.
State the recording, capture method and settings with any result. A clean sample does not prove performance in a noisy room or on another platform. Have a person verify consequential attributions before using them in minutes.
