transcription.py

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
import whisper



tokenizer = AutoTokenizer.from_pretrained("Bhuvana/t5-base-spellchecker")

model = AutoModelForSeq2SeqLM.from_pretrained("Bhuvana/t5-base-spellchecker")


def correct(inputs):
    input_ids = tokenizer.encode(inputs,return_tensors='pt')
    sample_output = model.generate(
        input_ids,
        do_sample=True,
        max_length=50,
        top_p=0.99,
        num_return_sequences=1
    )
    res = tokenizer.decode(sample_output[0], skip_special_tokens=True)
    return res

whisper_model = whisper.load_model("base")
def transcribe(audio_file):
    # Load audio and pad/trim it to fit 30 seconds
    audio = whisper.load_audio(audio_file)
    audio = whisper.pad_or_trim(audio)

    # Convert audio data to PyTorch tensor and float data type
    mel = torch.from_numpy(audio).float()

    # Make log-Mel spectrogram and move to the same device as the model
    mel = whisper.log_mel_spectrogram(mel).to(model.device)

    # Detect the spoken language
    _, probs = whisper_model.detect_language(mel)

    # Decode the audio
    options = whisper.DecodingOptions(fp16=False)
    result = whisper.decode(whisper_model, mel, options)
    result_text = result.text
    
    print('result_text:'+result_text)

    return correct(result_text)