pipeline.py

import numpy as np
from typing import Dict

from transformers import  Wav2Vec2Processor, Wav2Vec2ForCTC
from pyctcdecode import Alphabet, BeamSearchDecoderCTC

class PreTrainedPipeline():
    def __init__(self, path):
        """
        Initialize model
        """
        self.processor = Wav2Vec2Processor.from_pretrained(path)
        self.model = Wav2Vec2ForCTC.from_pretrained(path)
        vocab_list = list(self.processor.tokenizer.get_vocab().keys())

        # convert ctc blank character representation
        vocab_list[0] = ""

        # replace special characters
        vocab_list[1] = "⁇"
        vocab_list[2] = "⁇"
        vocab_list[3] = "⁇"

        # convert space character representation
        vocab_list[4] = " "

        alphabet = Alphabet.build_alphabet(vocab_list, ctc_token_idx=0)

        self.decoder = BeamSearchDecoderCTC(alphabet)
        self.sampling_rate = 16000


    def __call__(self, inputs)-> Dict[str, str]:
        """
        Args:
            inputs (:obj:`np.array`):
                The raw waveform of audio received. By default at 16KHz.
        Return:
            A :obj:`dict`:. The object return should be liked {"text": "XXX"} containing
            the detected text from the input audio.
        """
        input_values = self.processor(inputs, return_tensors="pt", sampling_rate=self.sampling_rate).input_values  # Batch size 1
        logits = self.model(input_values).logits.cpu().detach().numpy()[0]
        return {
            "text": self.decoder.decode(logits)
        }