I've been conducting a systematic evaluation of Sembly for our sales team's call review process, and a persistent technical issue is undermining its utility. In multiple recorded meetings with three distinct participants, Sembly's speaker diarization has incorrectly assigned a single speaker label (often "Speaker 2") to all three different voices. This collapses the conversation flow and makes the AI-generated insights and action items practically useless.
From my analysis, the problem isn't consistent across all recordings. It appears more frequently under these conditions:
* Meetings where participants join via phone lines (PSTN) rather than a unified VoIP client.
* Early-stage meetings where the system hasn't "learned" voices yet, though it should handle initial diarization.
* Recordings where background noise or varying audio levels are present, though not excessively so.
I have reviewed the technical prerequisites:
* Audio source is a high-quality conference bridge line, not a laptop microphone.
* Recording format is the recommended single-channel, WAV file.
* Each participant's audio is clear and distinguishable to a human listener.
My primary questions for the community are:
1. Has anyone identified a specific technical setup (codec, sample rate, etc.) that reliably triggers this bug?
2. Are there any documented workarounds, such as pre-processing audio files or a specific meeting initiation sequence, that force proper speaker separation?
3. From a TCO perspective, how are you mitigating this? Manual correction post-meeting negates the automation ROI.
I will share any findings from my continued tests. This is a critical flaw for any team considering Sembly for multi-party meeting analysis.
- Mark
independent eye
That PSTN angle is interesting. I've seen similar issues where separate phone lines get merged into a single audio stream by the conference bridge itself before the recording even happens. The diarization engine then has no way to split them.
Could you check the raw waveform in something like Audacity? If all three voices are on the same channel without clear pauses or level differences, it's a source problem, not a Sembly problem. You might need a different bridge setup or individual recordings per line.
Have you tried uploading a sample from a pure VoIP meeting as a control? It would isolate if the issue is truly format-related.
Pipeline Pilot
Your analysis on PSTN versus VoIP is likely on the right track. The conference bridge is often the culprit, outputting a mixed mono stream where diarization has no channel-level signal to separate speakers.
Have you checked if your bridge offers an "unmixed" or "individual stream" recording output? Some professional systems can record each line to a separate track in a multi-channel file. If Sembly supports multi-channel input, that would give the engine a direct mapping.
I'd test with a synthetic file: take a clean VoIP recording, duplicate it three times on separate channels in an audio editor, and upload it. If Sembly then labels three distinct speakers, you've confirmed the issue is with the audio source preprocessing, not the diarization model itself.