audio2text/a2t.py

import librosa
import numpy as np
from .init import pipe

TASK = "transcribe"
BATCH_SIZE = 8
LIMIT = 60

class A2T:
    def __init__(self, mic):
        self.mic = mic

    def __transcribe(self, inputs, task: str = None):
        if inputs is  None:
            print("Inputs None")

        transcribed_text = pipe(inputs, batch_size=BATCH_SIZE,)["text"]
        return transcribed_text

    def __preprocces(self, raw: np.ndarray, sampling_rate: int):
        chunk = raw.astype(np.float32) / 32678.0

        if sampling_rate != 16000:
            chunk = librosa.resample(chunk, orig_sr=sampling_rate, target_sr=16000)

        # chunk = chunk[:16000*LIMIT]
                
        return chunk
        
    def predict(self):
        try:
            if self.mic is not None:
                chunk = self.mic.get_array_of_samples()
                chunk = np.array(chunk, dtype=np.int16)
                sampling_rate = self.mic.frame_rate
                audio = self.__preprocces(raw=chunk, sampling_rate=sampling_rate)
                print(f"audio : {audio} \n shape : {audio.shape} \n max : {np.max(audio)}")
            else:
                raise Exception("please provide audio")

            if isinstance(audio , np.ndarray):
                return self.__transcribe(inputs=audio, task=TASK)
            else:
                raise Exception("Audio is not np array")
                
        except Exception as e:
            return f"Oops some kinda error : {e}"