refactor: enhance ASR model initialization and update translation configuration

app.py

@@ -16,7 +16,7 @@ from fastapi import FastAPI, UploadFile, File, HTTPException, Form
 from fastapi.responses import FileResponse
 from fastapi.middleware.cors import CORSMiddleware
 from starlette.background import BackgroundTask
-from transformers import pipeline, VitsModel, VitsTokenizer
+from transformers import pipeline, VitsModel, VitsTokenizer, AutoModelForSpeechSeq2Seq, AutoProcessor
 
 # External service imports
 from google import genai
@@ -36,12 +36,23 @@ logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(
 async def lifespan(app: FastAPI):
     # Load models once at startup and store in app.state
     device = "cuda:0" if torch.cuda.is_available() else "cpu"
+    torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
     model_id = "rbcurzon/whisper-large-v3-turbo"
+    
+    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
+    )
+    model.to(device)
+    
+    processor = AutoProcessor.from_pretrained(model_id)
+    
     app.state.pipe = pipeline(
         "automatic-speech-recognition",
-        model=model_id,
-        chunk_length_s=30,
-        device=device
+        model=model,
+        tokenizer=processor.tokenizer,
+        feature_extractor=processor.feature_extractor,
+        torch_dtype=torch_dtype,
+        device=device,
     )
     app.state.vad_model = load_silero_vad()
     app.state.client = genai.Client(api_key=os.environ.get("GENAI_API_KEY"))
@@ -81,6 +92,7 @@ def translate(text, srcLang, tgtLang):
         config=types.GenerateContentConfig(
             system_instruction=f"You are an expert translator. Your task is to translate from {srcLang} to {tgtLang}. You must provide ONLY the translated text. Do not include any explanations, additional commentary, or conversational language. Just the translated text.",
             thinking_config=types.ThinkingConfig(thinking_budget=0),  # Disables thinking
+            temperature=0.6,
         )
     )
     return response.text
@@ -116,10 +128,11 @@ async def translate_audio(
             "num_beams": 1,
             "condition_on_prev_tokens": False,
             "compression_ratio_threshold": 1.35,
-            "temperature": 0.0,
+            "temperature": 0.0, # reduce temperature for more deterministic output
             "logprob_threshold": -1.0,
             "no_speech_threshold": 0.6,
             "return_timestamps": True,
+            "language": "tl"
         }
 
         temp_file = remove_silence(file.filename)