sdpa

app.py

@@ -35,8 +35,8 @@ pipe = pipeline(
     model="openai/whisper-large-v3-turbo",
     torch_dtype=torch.float16,
     device="cuda",
-    model_kwargs={"attn_implementation": "flash_attention_2"},
-    return_timestamps="word",
+    model_kwargs={"attn_implementation": "sdpa"},#flash_attention_2
+    return_timestamps=True,
 )
 
 class WhisperTranscriber:
@@ -115,7 +115,7 @@ class WhisperTranscriber:
         # Transcribe with timestamps
         result = self.pipe(
             audio_path,
-            return_timestamps="word",
+            return_timestamps=True,
             generate_kwargs=generate_kwargs,
             chunk_length_s=30,
             batch_size=128,