audio2text/a2t.py

import numpy as np

import librosa
import torch
from .init import processor, model

LIMIT = 90 # limit 90 seconds

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

    def __preprocces(self, audio, frame_rate):
        try:
            print("Audio before : ", audio)
            audio = audio / 32678.0
            print("Audio div : ", audio)

            if len(audio.shape) > 1: 
                audio = librosa.to_mono(audio.T)
                
            print("Audio mono : ", audio)
    
            if frame_rate != 16_000:
                audio = librosa.resample(audio, orig_sr=frame_rate, target_sr=16000)

            print("Audio resample : ", audio)
    
            audio = audio[:16_000*LIMIT]

            print("Audio cut : ", audio)
    
            audio = torch.tensor(audio)

            print("Audio torch : ", audio)
            return audio
        except Exception as e:
            print("Preprocces error", e)
            return None
        
    def predict(self):
        if self.mic is not None:
            audio = self.mic
            # frame_rate = self.mic.frame_rate
        else:
            return "please provide audio"
    
        try:
            forced_decoder_ids = processor.get_decoder_prompt_ids(language="english", task="transcribe")
            # audio = self.__preprocces(audio=audio, frame_rate=frame_rate)
            inputs = processor(audio=audio, sampling_rate=16000, return_tensors="pt")
            predicted_ids = model.generate(**inputs, max_length=400, forced_decoder_ids=forced_decoder_ids)
            transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
            return transcription[0]
        except Exception as e:
            print("Predict error", e)
            return "Oops some kinda error"