| --- |
| title: Diarized Speaker Segments Community-1 |
| emoji: 🎙️ |
| colorFrom: green |
| colorTo: blue |
| sdk: docker |
| app_port: 7860 |
| suggested_hardware: t4-small |
| --- |
| |
| # Diarized Speaker Segments Community-1 |
|
|
| This Space uses **repo-style transcription logic** from your attached codebase plus **pyannote/speaker-diarization-community-1**. |
|
|
| ## What changed |
| - transcription now uses the attached repo's logic: |
| - custom speech-window detection |
| - ffmpeg audio enhancement |
| - repo-style Hindi/Hinglish/English prompt |
| - repo thresholds and dedupe flow |
| - repo segment splitting by words |
| - diarization remains `community-1` |
| - cleanup remains exactly: |
| - merge **only adjacent same-speaker** segments |
| - otherwise do not touch |
|
|
| ## Notes |
| - default ASR model is **medium** |
| - `large-v3` is available as a dropdown for evaluation |
| - default language is **hi** because the attached repo logic is Hindi-biased for Hinglish handling |
| - you can still test `auto` and `en` |
|
|
