Speaker Diarization: Know Who Said What

Speaker diarization is the process of partitioning an audio recording by speaker - answering "who spoke when" - so a transcript reads as attributed dialogue instead of one undifferentiated block of text. It groups speech segments that belong to the same voice and labels each one (Speaker 1, Speaker 2, or a real name). It is distinct from speech recognition, which turns audio into words; diarization adds the layer of who said each of those words.

Meetily Pro identifies each speaker and labels every segment of the transcript with who said it - live as you record and on audio you import or re-transcribe. It runs on-device, so no audio is uploaded to label speakers. Rename a speaker once and the name updates across the entire transcript, and you can merge duplicates if the same person was split into two. Diarization is part of Meetily Pro (since version 1.8.1); the app is open-core with 25K+ GitHub stars.

Speaker diarization separates a recording into distinct speakers without knowing who they are - it produces anonymous labels like Speaker 1 and Speaker 2. Speaker identification (or recognition) goes a step further and matches a voice to a known identity. In practice, product UIs use the terms interchangeably: Meetily diarizes each meeting into separate speakers, then lets you name them once so the label carries across the transcript. No enrolled voiceprint database is required.

In Meetily, diarization runs entirely on-device. Audio never leaves your machine to be split by speaker - the same on-device pipeline that transcribes with Whisper or Parakeet also attributes the speech. This matters for confidential meetings: legal, healthcare, HR, and executive conversations stay on your own hardware. For summaries you separately choose a local model or your own API key (transcript text only, opt-in per meeting).

Yes. Meetily attributes speakers live as you record, and also on audio files you import or on past meetings you re-transcribe. Import a recording from any source, run it through Meetily, and the resulting transcript is labeled by speaker. Re-transcribing an older meeting applies diarization to it as well, so your back catalog can gain speaker labels.

Open-source diarization usually means assembling a pipeline yourself - for example OpenAI Whisper for transcription plus pyannote.audio or NVIDIA NeMo for speaker segmentation - which takes setup, tuning, and GPU know-how. Meetily takes a packaged approach: its open-core app pairs on-device Whisper/Parakeet transcription with built-in diarization in the Pro edition, so you get who-said-what labeling without wiring a pipeline together. Everything runs locally on Windows or macOS.

Speaker diarization is a Meetily Pro feature (introduced in Pro 1.8.1). The free, MIT-licensed Community Edition includes local transcription and AI summaries, but not diarization. Pro is $10/user/month billed annually (early bird) and adds speaker identification, more accurate models, custom summary templates, and priority support. A 14-day trial is available with no card required.

Diarization accuracy depends on audio quality, how many speakers there are, how much they overlap, and microphone setup - clean, low-overlap audio attributes more reliably than a crowded room on one mic. Meetily lets you correct attribution: rename a speaker once and it propagates, and merge duplicates when one person is split across two labels. That keeps the final transcript accurate even when the automatic pass is imperfect.

Meetily captures audio from your own device rather than joining the call as a bot, so speaker diarization works the same across Zoom, Microsoft Teams, Google Meet, and any other platform. There is no per-platform integration to configure. It works on both live calls and imported recordings, on Windows and macOS.