Update README.md

README.md

@@ -10,24 +10,22 @@ suggested_hardware: t4-small
 
 # Diarized Speaker Segments Community-1
 
-This Space uses **pyannote/speaker-diarization-community-1** and `faster-whisper`.
+This Space uses **repo-style transcription logic** from your attached codebase plus **pyannote/speaker-diarization-community-1**.
 
-## Important
-- **Default ASR model is `medium`**
-- `large-v3` is available as a dropdown for evaluation
-- **Default language is `auto`**
-- For **Hinglish / mixed Hindi-English audio**, keep language on **`auto`**
-- Do **not** force `hi` or `en` unless you are explicitly testing that behavior
-
-## Cleanup behavior
-This version only does one cleanup step after speaker assignment:
+## 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
 
-- if **adjacent transcript segments have the same speaker**, merge them
-- preserve the earliest start and latest end timestamp
-- do **not** do extra smoothing beyond that
-
-## Runtime behavior
-- ASR uses GPU when available
-- diarization uses GPU when available
-- ASR and diarization run sequentially
-- diarization is fed an in-memory waveform dict to avoid file-decoding issues
+## 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`