ffmpeg_handler.py

from typing import Dict, Any, List
from transformers import WhisperForConditionalGeneration, AutoProcessor, WhisperTokenizer, WhisperProcessor, pipeline, WhisperFeatureExtractor
import torch
from transformers.pipelines.audio_utils import ffmpeg_read
#import io


#device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


class EndpointHandler:
    def __init__(self, path=""):
        #tokenizer = WhisperTokenizer.from_pretrained('openai/whisper-large', language="korean", task='transcribe')        
        #model = WhisperForConditionalGeneration.from_pretrained(path)
        #self.tokenizer = WhisperTokenizer.from_pretrained(path)
        #self.processor = WhisperProcessor.from_pretrained(path, language="korean", task='transcribe')
        #processor = AutoProcessor.from_pretrained(path)
        #self.pipe = pipeline("automatic-speech-recognition", model=model, tokenizer=processor.feature_extractor, feature_extractor=processor.feature_extractor)
        #feature_extractor = WhisperFeatureExtractor.from_pretrained('openai/whisper-large')
        self.pipe = pipeline(task='automatic-speech-recognition', model=path)
        

        
        # Move model to device
#        self.model.to(device)
        
    def __call__(self, data: Any) -> List[Dict[str, str]]:
        print('==========NEW PROCESS=========')

        inputs = data.pop("inputs", data)
        audio_nparray = ffmpeg_read(inputs, 16000)
        audio_tensor= torch.from_numpy(audio_nparray)

        
        transcribe = self.pipe
        transcribe.model.config.forced_decoder_ids = transcribe.tokenizer.get_decoder_prompt_ids(language="ko", task="transcribe")
        result = transcribe(audio_tensor)
        

        return result