handler.py

from typing import Dict
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from transformers.pipelines.audio import AudioClassificationPipeline
from datasets import load_dataset
import torch

SAMPLE_RATE = 16000

class EndpointHandler():
    def __init__(self, path=""):
        # load the model
        self.processor = WhisperProcessor.from_pretrained("openai/whisper-large-v2")
        self.model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v2")
        self.classifier = AudioClassificationPipeline(model=self.model, processor=self.processor, device=0)
        self.forced_decoder_ids = self.processor.get_decoder_prompt_ids(language="Danish", task="transcribe")

    def __call__(self, data: Dict[str, bytes]) -> Dict[str, str]:
        """
        Args:
            data (:obj:):
                includes the deserialized audio file as bytes
        Return:
            A :obj:`dict`:. base64 encoded image
        """
        # process input
        inputs = data.pop("inputs", data)
        audio_nparray = ffmpeg_read(inputs, sample_rate=SAMPLE_RATE)
        audio_tensor= torch.from_numpy(audio_nparray)

        # run inference pipeline
        result = self.classifier(audio_nparray)

        # postprocess the prediction
        return {"txt": result[0]["transcription"]}