# Reference-based review

The model detected two speakers, but speaker assignments and short replies still need review.

Reference-based comparison against the AMI manual word annotations v1.6.2, including speaker IDs and original source times. Not a listening audit.

- The first model segment (“Because, you know…booking system”) corresponds to source speaker A, while the next (“Well, we think…”) corresponds to B. Both are labeled SPEAKER_00, so this boundary does not preserve the two source speakers.
- The 38.30–43.35s account-related statements are from reference speaker B, but the model uses SPEAKER_01, also used for speaker A’s earlier question. The labels are not consistently assigned.
- The reference’s “the two key holders are here” is not present in the model transcript, and opening/overlapping short replies are omitted.
- The person-name reference “Marloes” appears as “Marlos”; “Caroline” is retained and J_P_ is rendered JP.
- At 58.04–60.92s “Because then they learn the first time” is assigned SPEAKER_00 even though the corresponding source words are annotated as A.

## Limits

- These are two short, selected excerpts from real non-scenario multi-person meetings, using the official mixed headset recording. They are not privately recorded one-person/two-person sessions.
- The reference contains false starts, fillers and overlapping speech. Word-count differences are not an error rate, and no WER/DER or general accuracy claim is made.
- Timestamps are model output, not independently listened-to alignment measurements. All normalized timing statuses remain unknown, even where numeric times exist.
- No output text, speaker label or timestamp was manually corrected. Original provider output and normalized original exports are preserved.
- This direct-provider test uses the unchanged production input adapter, pinned model, normalizer and exporters. It does not test application login, credits, job database, webhook handling or browser processing.