switch to bfloat16

app.py

@@ -42,7 +42,8 @@ model = AutoModelForSpeechSeq2Seq.from_pretrained(
     use_cache=False,
     device_map="auto",
     low_cpu_mem_usage=True,
-    attn_implementation="flash_attention_2"
+    attn_implementation="flash_attention_2",
+    torch_dtype=torch.bfloat16
 )
 
 # Flash Attention setup for memory and speed optimization if supported
@@ -64,7 +65,7 @@ pipe = pipeline(
     model=model,
     tokenizer=tokenizer,
     feature_extractor=feature_extractor,
-    chunk_length_s=900,  # Increased to 15 minutes
+    chunk_length_s=30,  # 30 seconds
 )
 
 def reset_and_update_dataset(new_data):