Public demo

Speaker Diarization Demo

Upload a short audio or video recording and compare how Azure AI Speech and GPT-4o label speakers in the transcript.

What you can do

  • Transcribe an uploaded audio or video file using Japanese or English (US) speech recognition.
  • Compare Azure AI Speech and GPT-4o diarization results.
  • Review timestamped speaker turns, copy the transcript, or download JSON.

Learn more

Read how speaker diarization differs from speaker recognition in Azure AI Speech.

How to use it

  1. Select a supported audio or video file.
  2. Select the recognition language: Japanese or English (US).
  3. For Azure AI Speech, set the expected maximum number of speakers to a value from 2 to 35. GPT-4o detects the number automatically.
  4. Choose Azure AI Speech or GPT-4o, then wait for security scanning, queueing, and analysis to complete.
  5. Review the result, then copy the transcript or download the JSON if needed.

Public demo limits

File size
Up to 50 MiB
Duration
Up to 5 minutes
Accepted formats
WAV, MP3, M4A, MP4, MOV, OGG, OPUS, FLAC, WMA, AAC, AMR, WebM
Per session
Up to 5 uploads accepted per UTC day
Per network identifier
Up to 30 uploads accepted per UTC day
Concurrent jobs
1 per session; 3 per network identifier
Shared capacity
150 minutes of audio per UTC day and 3,000 minutes per UTC month

Limits may change based on usage, service capacity, and operating cost.

Accuracy and appropriate use

  • Speaker labels are anonymous and apply only within a single result. They do not identify real people or indicate that a speaker is the same person across different files or engines.
  • Overlapping speech, brief responses, background noise, recording quality, and microphone distance can reduce accuracy.
  • Results may contain errors. Do not use them as evidence or for identity verification, personnel evaluations, or other high-impact decisions.
  • Obtain the necessary consent before uploading another person’s voice.

Data handling

  • Uploaded audio is sent to Azure AI Services after malware scanning.
  • Input audio is deleted after processing succeeds or definitively fails. Any orphaned blobs are deleted after one day under the storage lifecycle policy.
  • Private result JSON is retained for up to one hour; job metadata without transcript text is retained for up to 24 hours.
  • Audio, transcript text, original filenames, raw IP addresses, cookies, and tokens are not written to application logs.

How the public endpoint is protected

  • The demo is exposed only through the shared memobog.net Azure Front Door endpoint. Direct access to its web and API origins is denied.
  • Azure Web Application Firewall inspects requests and applies an upload rate limit before traffic reaches the demo.
  • The API origin uses Azure Front Door Premium Private Link. Browser responses also enforce Content Security Policy and Permissions Policy headers.