Update handler.py

handler.py

@@ -1,34 +1,31 @@
 from typing import Dict
 from faster_whisper import WhisperModel
 import io
-import re
 
 class EndpointHandler:
     def __init__(self, model_dir=None):
-        # The compute_type is set to "float16" for efficient GPU computation
-        # For "int8" computation on CPU, the compute_type would be "int8"
-        compute_type = "float16"
-
-        # Initialize WhisperModel with large-v2 model size and specified compute_type
-        model_size = "large-v2" if model_dir is None else model_dir
-        self.model = WhisperModel(model_size, device="cuda", compute_type=compute_type)
+        # Set model size, assuming installation has been done with appropriate model files and setup
+        model_size = "medium" if model_dir is None else model_dir
+        # Change to 'cuda' to use the GPU, and set compute_type for faster computation
+        self.model = WhisperModel(model_size, device="cpu", compute_type="int8")
 
     def __call__(self, data: Dict) -> Dict[str, str]:
+        # Process the input data expected to be in 'inputs' key containing audio file bytes
         audio_bytes = data["inputs"]
+        
+        # Convert bytes to a file-like object
         audio_file = io.BytesIO(audio_bytes)
-
-        # Transcribe audio file with a smaller beam size for faster inference
-        # Note: Adjust beam_size based on desired accuracy vs speed trade-off
-        beam_size = 1
-        segments, info = self.model.transcribe(audio_file, beam_size=beam_size)
-
-        # Aggregate transcribed text and remove any extra spaces
+        
+        # Perform transcription using the model
+        segments, info = self.model.transcribe(audio_file)
+        
+        # Compile the results into a text string and extract language information
+        # Strip whitespace from each segment before joining them
         text = " ".join(segment.text.strip() for segment in segments)
-        text = re.sub(' +', ' ', text)
-
         language_code = info.language
         language_prob = info.language_probability
 
+        # Compile the response dictionary
         result = {
             "text": text,
             "language": language_code,