Fix audio duration detection using librosa for accurate duration measurement

src/api/transcriptions.py

@@ -17,6 +17,7 @@ from src.services.whisper_service import (
 from src.services.stats_service import log_usage
 from src.database.db import get_db
 from sqlalchemy.orm import Session
+import librosa
 
 router = APIRouter()
 
@@ -202,6 +203,13 @@ async def create_transcription(
         
         file_size = len(content)
         
+        # Get audio duration using librosa (more reliable than whisper's duration)
+        try:
+            audio_duration = librosa.get_duration(path=temp_path)
+        except Exception as e:
+            print(f"Warning: Could not get audio duration: {e}")
+            audio_duration = None
+        
         # Transcribe
         include_word_timestamps = timestamp_granularities and "word" in timestamp_granularities
         
@@ -213,13 +221,16 @@ async def create_transcription(
         
         processing_time = int((time.time() - start_time) * 1000)
         
+        # Use audio_duration if available, otherwise fall back to whisper's duration
+        duration_to_log = audio_duration if audio_duration else result.get("duration")
+        
         # Log usage
         await log_usage(
             db=db,
             api_key=api_key,
             endpoint="/v1/audio/transcriptions",
             file_size=file_size,
-            duration=result.get("duration"),
+            duration=duration_to_log,
             processing_time=processing_time,
             model=settings.whisper_model,
             status="success"