Skip to workspace
PixelOkay
Browser-local audio
Tools
AUDIO INTELLIGENCE STUDIO V4

AI Speaker Diarization

Explore speaker turns, build a labelled transcript, and understand who spoke when. AI-assisted analysis with complete human control.

◎
Your recording stays with youAudio is decoded locally. Free models download only after you approve model setup; recording content is not uploaded.
WelcomeOpen a recording to begin. You can also edit speaker intervals manually without downloading any AI model.
RECORDING

Speaker timeline

Start with a conversation

Drag an audio file here, or open one from your device.

Speaker labels
0
Speech intervals
0
Recording length
—
ANNOTATION WORKSPACE

Speaker turns 0

No speaker turns yet. Run the analysis or add a manual interval.
No intervals

How speaker diarization works

01

Open your audio

Load a WAV, MP3 or other format your browser can decode. No recording is uploaded.

02

Analyze, then verify

Speaker AI labels local voice activity. Optional English Whisper transcription is matched to intervals where timestamps exist.

03

Correct and export

Adjust speakers, times, or words. Export subtitles, a diarization timeline or an editable JSON project.

Important: Speaker IDs are provisional, especially across processing chunks. This is not personal identity verification, forensic speaker attribution or guaranteed word-level transcription. Model downloads require network access on first use; audio inference happens on your device. Long recordings remain available for manual annotation.

LOCAL AI SETUP

Set up AI on this device

First, load the speaker models on your device

Speech segmentation and optional English transcription run in this browser. Free model weights and runtime code may download from Hugging Face and a CDN after you approve. Files already cached by compatible PixelOkay tools may be reused. No audio is uploaded for inference.

  1. 1. Check device
  2. 2. Download / cache
  3. 3. Set up AI
  4. 4. Analyze audio

Pyannote segmentation 3.0 (ONNX + WASM); optional Whisper tiny.en (ONNX + WASM). Both are real audio models, not chat models. Exact download sizes and RAM needs vary and are not estimated here.

Device & model resources

Approximate where supported
Device RAM / model memoryChecking device information…

Live system RAM and model memory aren't reliably exposed by the browser.

Browser website storageChecking storage…

Site-wide estimate; includes other PixelOkay tools. This is not RAM.

ONE ACTION · ALL RECOGNIZED AI MODEL DOWNLOADS

Free AI model storage across PixelOkay

Scan this website origin for WebLLM and Transformers.js model caches, plus recognizable model files in other cache buckets. Includes models from other PixelOkay tools, not only this one. Unrelated website data and databases are never automatically cleared.

Inspect to see recognized model files; this cannot identify every runtime-managed cache.

Unknown browser-managed IndexedDB / OPFS model storage is not erased. Close other AI tools before deletion. Saved chats, files, and projects are not targets.

Having trouble downloading or loading AI?

Try refreshing Storage & Cache, checking your network, then Retry Setup. Cache deletion is optional and only for positively recognized model assets; it cannot fix all runtime errors.

Clearing all website data is a last resort and may permanently remove saved work from multiple PixelOkay tools. Export your projects first.

Your audio can be annotated manually without an AI model.
Currently loadedNone

No model in worker memory

The Pyannote model predicts local speaker slots; it does not identify people. Returning speakers may receive different labels. English transcription is approximate and must be reviewed.

Delete recognized AI model files?

Only recognized model files will be targeted.

Model files used by other PixelOkay tools may need to download again. Close other AI tabs first. Unknown site storage is excluded.

Using AI Speaker Diarization

Import a supported recording and annotate turns manually or request optional AI analysis. Review every speaker label and unassigned transcript before editing and exporting. Save a project with the matching original recording.

Segmentation slots are local to analysis windows, not persistent identities. Returning speakers can receive new labels; labels/confidence require human review. Optional models/runtime download only after approval in AI Model Setup; AI analysis has a shorter duration budget than import. No speaker recognition, forensic identity proof or ground-truth transcript is supplied. JSON excludes audio; file/memory/label limits apply.

References: Segmentation model · English transcription model

Label provisional speaker turns and verify the transcript

Import a supported recording and annotate turns manually, or approve optional segmentation and English transcription models in Model Setup before analysis. Review the time ranges, speaker labels and unassigned words against the audio before exporting captions, RTTM or a project.

Example workflow

  1. Open a short conversation and listen through once to learn the voices and turn order.
  2. Run optional analysis if useful, then review each provisional turn and any unassigned transcript instead of accepting model labels automatically.
  3. Correct speakers and timings, then export the format needed for captioning or downstream review. Keep the original recording with a saved project.

Authored example: In a two-person interview, inspect a short exchange where one person briefly interrupts. Check the overlap against the audio and relabel the turn manually; the same person may receive another slot in a later analysis window.

Limits and model handling: Optional Pyannote segmentation and English Whisper assets download only after approval and run as local models. Labels describe provisional activity, not verified people or persistent identity. Returning speakers can receive new labels, and transcription/confidence need human review. Model analysis has a shorter duration budget than import. Project JSON excludes the audio.

AI Speaker Diarization questions

Does a speaker label identify a real person across recordings?

No. Labels are provisional analysis slots and are not identity recognition or forensic proof; the same speaker can receive a different label in another window.

Can the transcript be used without proofreading?

No. Review words, unassigned text, timings and overlaps against the original recording before exporting.

Related tools