change transcribe mechanism when uploading audio

Dockerfile

@@ -29,4 +29,4 @@ RUN pip install --no-cache-dir --upgrade -r requirements.txt
 
 COPY --chown=user . /app
 # CMD ["gunicorn", "-b", "0.0.0.0:7860", "app:app"]
-CMD ["gunicorn", "--timeout", "120", "--workers", "2", "-b", "0.0.0.0:7860", "app:app"]
+CMD ["gunicorn", "--timeout", "120", "-b", "0.0.0.0:7860", "app:app"]

app.py

@@ -9,7 +9,6 @@ import requests
 from tqdm import tqdm
 from transformers import BertTokenizer, AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
 from model.multi_class_model import MultiClassModel  # Adjust if needed
-from optimum.intel import OVModelForSpeechSeq2Seq
 
 app = Flask(__name__)
 
@@ -50,48 +49,35 @@ model = MultiClassModel.load_from_checkpoint(
 )
 model.eval()
 
-# === OPENVINO WHISPER INIT ===
-def init_whisper_openvino():
-    device = "cpu"  # Force CPU for OpenVINO
-    model_id = "openai/whisper-large-v3"
-
-    # Load OpenVINO-optimized model
-    ov_model = OVModelForSpeechSeq2Seq.from_pretrained(
-        model_id,
-        export=True,  # Auto-convert to OpenVINO format
-        compile=False,
-        trust_remote_code=True,
-    )
-
-    # Compile for specific input shapes
-    ov_model.compile(use_auto=True)
-
-    processor = AutoProcessor.from_pretrained(model_id)
-
-    return pipeline(
-        "automatic-speech-recognition",
-        model=ov_model,
-        feature_extractor=processor.feature_extractor,
-        tokenizer=processor.tokenizer,
-        max_new_tokens=128,
-        chunk_length_s=30,
-        batch_size=1,  # Optimized for OpenVINO on CPU
-        device=device,
-        torch_dtype="float32",
-    )
-
-# Initialize once at startup
-whisper_pipe = init_whisper_openvino()
+# === INITIAL SETUP: Whisper Pipeline ===
+# https://huggingface.co/openai/whisper-large-v3
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+model_id = "openai/whisper-large-v3"
+
+whisper_model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
+)
+whisper_model.to(device)
+
+processor = AutoProcessor.from_pretrained(model_id)
+
+pipe = pipeline(
+    "automatic-speech-recognition",
+    model=whisper_model,
+    tokenizer=processor.tokenizer,
+    feature_extractor=processor.feature_extractor,
+    chunk_length_s=10,
+    batch_size=4,  # batch size for inference - set based on your device
+    torch_dtype=torch_dtype,
+    device=device,
+)
+
 
 def whisper_api(temp_audio_path):
-    result = whisper_pipe(
-        temp_audio_path,
-        return_timestamps=False,
-        generate_kwargs={
-            "language": "indonesian",
-            "task": "transcribe",
-        }
-    )
+    result = pipe(temp_audio_path, return_timestamps=False, generate_kwargs={"language": "indonesian"})
+    print(result["text"])
     return result
 
 

requirements.txt

@@ -12,7 +12,6 @@ setuptools-rust
 # ffmpeg
 # ffmpeg-python
 # imageio[ffmpeg]
-torchaudio
 accelerate
 pytorch-lightning==2.2.1
 lightning==2.4.0