Skip to content

Multi-speaker file imports misattribute speakers even without channel collapse #6

Description

@predict-woo

Summary

Imported recordings with multiple speakers can produce significant speaker misattribution even when the source does not contain isolated channels that could be lost during downmixing.

Steps to reproduce

  1. Import a plain stereo recording containing two speakers.
  2. Let import transcription and diarization complete.
  3. Compare the assigned speaker IDs with the recording.

Actual behavior

Speech is attributed to the wrong speaker. A one-speaker control imported correctly, while a two-speaker plain-stereo file reproduced the problem. A separate four-person import folded much of one person's speech into the other identities.

Expected behavior

File-import diarization should distinguish speakers with accuracy comparable to live-recorded audio.

Notes

This is separate from the confirmed channel-collapse issue: #5. The two-speaker reproduction had no isolated per-speaker tracks to lose, so unconditional downmixing does not fully explain it.

Reported by Joey Virrueta.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions