Research conducted October 2026 from official documentation, research papers and legal references. Details may change.
Short answer: speaker diarization answers "who spoke when." It splits a recording into segments and groups them by voice, so a transcript can say Speaker 1 and Speaker 2. It's a separate step from turning speech into words [1].
Diarization vs. transcription
Transcription (speech-to-text) produces the words. Diarization produces a timeline of speaker turns. A meeting app that shows labeled speakers is running both and lining them up by timestamp. Speech recognizers like Whisper focus on the words and don't label speakers on their own [6].
How a typical diarization pipeline works
Most systems follow the same stages, described in detail in Park et al.'s review [1]:
- Voice activity detection. Find the parts of the audio that contain speech [5].
- Segmentation. Cut speech into short pieces, ideally each from a single speaker, and detect speaker changes.
- Speaker embeddings. Turn each piece into a vector, a "voiceprint," using a neural network trained to tell voices apart.
- Clustering. Group similar vectors. Each cluster becomes one speaker. The system often also has to guess how many speakers there are.
- Overlap handling. Newer neural models can label moments where two people talk at once [1][3].
Open-source toolkits like pyannote.audio package these steps into pretrained pipelines [2][3].
How accuracy is measured
Researchers use diarization error rate (DER): the share of time that is missed speech, false alarms, or assigned to the wrong speaker [1]. Benchmarks like the DIHARD challenge deliberately use hard audio such as restaurants, clinical interviews and web video, where error rates are much higher than in clean studio recordings [4].
Why it's hard in real meetings
- Overlapping speech is common in lively discussions [4].
- Similar voices produce embeddings that cluster together.
- One far-away microphone (a phone on a table) captures echo and uneven volume.
- Unknown speaker count. The system must decide whether two segments are one person or two.
Do you need speaker labels?
For interviews and legal or HR notes, labeled turns are valuable. For many meetings, what matters most is decisions, action items and a searchable transcript. Keepsay focuses on that: on-device transcription with no account for on-device use [7], and it adds recording modes and a To-dos tab. Keepsay doesn't label speakers today.
Tips if you rely on diarization
- Use a separate microphone per person when you can.
- Ask people not to talk over each other on important points.
- Spot-check speaker labels before sharing a transcript; clustering mistakes swap whole sections.
Related: iPhone meeting recorder apps in 2026
Sources & methodology
Methodology. In October 2026 we read the official documentation, papers, repositories and legal references listed below, and checked that every link loaded. We did not run benchmark tests. Keepsay facts come from its App Store listing, website and app code; Keepsay supports 12 recording languages.