app.py

# imports for
from datasets import Audio, load_dataset, Dataset
import torch
from transformers import pipeline

from pathlib import Path

# for the UI by Gradio
import gradio as gr
import pandas as pd


# initializing the values for device
if torch.cuda.is_available():
    DEVICE = "cuda:0"
    TORCH_DTYPE = torch.float16
else:
    DEVICE = "cpu"
    TORCH_DTYPE = torch.float32


MODEL_NAME = 'openai/whisper-large-v3'

pipe = pipeline(
    "automatic-speech-recognition",
    model=MODEL_NAME,
    torch_dtype=TORCH_DTYPE,
    device=DEVICE,
)




def convert_audio_2_array(files):
  file_paths = files
  complaint_data = Dataset.from_dict({		
      "audio" :  [file_paths] # path to be changed based on the path
      }).cast_column("audio",Audio(sampling_rate=16000)) 
  file_name = Path(file_paths).name
  print(f"file_paths: \n {file_paths} and \n file_name: {file_name}and \n complaint_data : \n {complaint_data} ")
  return file_name, complaint_data


def v2t_convertor(files):
  file_name,inputs_dict =  convert_audio_2_array(files)
  
  input = inputs_dict[0] # selecting only one input
  org_complain_dict = pipe(input["audio"].copy(), max_new_tokens=256, generate_kwargs={"task": "transcribe"})
  print('f{org_complain_dict}')
  org_complain = org_complain_dict['text']

  # lang_token = pipe.model.generate(input, max_new_tokens=1)[0,1]
  # language_code = pipe.tokenizer.decode(lang_token)
  language_code = "hi" # hard coded for the time being

  eng_complain = pipe(input["audio"].copy(), max_new_tokens=256, generate_kwargs={"task": "translate"})['text']

  print(f"org_complain: \n {org_complain} \nand\n eng_complain:\n {eng_complain} \n language_code: {language_code}")

  return [[file_name, org_complain, eng_complain, language_code]]



def upload_file(files): # the actual translation should happen here
    """
    takes the file that comes from the UI and converts it to the respective
    format to be sent to the model for transcription
    """
    # Define the column names
    columns = ["audio_id", "transcribed_text( in org lang )", "transcribed_text( in eng )", "language"]

    # Define the data as lists. Here the data would be sent in the form of single data fro mthe data

    # data = [ ["ca_1.wav", "बिना किसी पूर्व सूचना के विलंबित या रद्द की गई ट्रिनिक", "without any prior information or any delay or delay in the train journey", "hindi"]]

    data = v2t_convertor(files)

    # Create the DataFrameoutputs
    if data is not None:
        df = pd.DataFrame(data, columns=columns)
    else:
        raise ValueError("Data is None. Cannot create DataFrame.")
    return df


with gr.Blocks() as demo:
  gr.Markdown( "# V2T Engine"  )
  with gr.Accordion("Steps to run the App"):
    gr.Markdown("1. Click the \"Click to Upload a File\" to get the dialog box to browse your local path." )
    gr.Markdown("2. The audio will get uploaded, transcribed into the ")
    gr.Markdown("3. ")

  upload_button = gr.UploadButton("Click to Upload a File", file_types=["audio"], file_count="single")
  df_output = gr.Dataframe(
            headers=["audio_id", "transcribed_text( in eng )", "transcribed_text( in org lang )", "language"],
            datatype=["str", "str", "str"],
            row_count=1,
            col_count=(4, "fixed"),
            wrap=True
    )
  upload_button.upload(upload_file, upload_button, df_output, show_progress = True) # upload the audio file and and sends to the upload function



if __name__ == "__main__":
    demo.launch(debug=True)