handler.py

from typing import Dict
from faster_whisper import WhisperModel, Streaming
import io
import re

class EndpointHandler:
    def __init__(self, model_dir=None):
        # Set model size, assuming installation has been done with appropriate model files and setup
        model_size = "large-v2" 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="cuda", compute_type="float16")

    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)
        
        # Enable VAD and perform transcription using the model with a reduced beam size
        streaming = Streaming(device="cuda", compute_type="float16", vad=True)
        segments, info = streaming.transcribe(audio_file, beam_size=1)
        
        # Compile the results into a text string and extract language information
        # Strip leading and trailing whitespace and replace multiple spaces with a single space
        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,
            "language_probability": language_prob
        }
        
        return result