app.py

import gradio as gr
import librosa
import torch
import torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import numpy as np

processor = Wav2Vec2Processor.from_pretrained("maher13/arabic-iti")
model = Wav2Vec2ForCTC.from_pretrained("maher13/arabic-iti").eval()

def asr_transcript(audio_file, audio_file2):
    transcript = ""

    if audio_file : 
      wav, sr = librosa.load(audio_file.name, sr=16000)
 
      input_values = processor(wav, sampling_rate=16000, return_tensors="pt", padding=True).input_values
      logits = model(input_values).logits
  
      with torch.no_grad():
        predicted_ids = torch.argmax(logits, dim=-1)
      predicted_ids[predicted_ids == -100] = processor.tokenizer.pad_token_id
      transcription1 = processor.tokenizer.batch_decode(predicted_ids)[0]
    else:
      transcription1 = "N/A"
    
    if audio_file2 : 
      wav, sr = librosa.load(audio_file2.name, sr=16000)

      input_values = processor(wav, sampling_rate=16000, return_tensors="pt", padding=True).input_values
      logits = model(input_values).logits
  
      with torch.no_grad():
        predicted_ids = torch.argmax(logits, dim=-1)
      predicted_ids[predicted_ids == -100] = processor.tokenizer.pad_token_id
      transcription2 = processor.tokenizer.batch_decode(predicted_ids)[0]
    else :
      transcription2 = "N/A"

    return transcription1, transcription2

gradio_ui = gr.Interface(
    fn=asr_transcript,
    title="Speech to Text Graduation project \n sponsored by TensorGraph",
    inputs=
    [
    gr.inputs.Audio(source = 'microphone', type="file", optional = True),
    gr.inputs.Audio(source = 'upload', type="file", optional = True) 
    ],
    outputs=[
             gr.outputs.Textbox(label="Auto-Transcript"),
             gr.outputs.Textbox(label="Auto-Transcript")
             ],
)



#gradio_ui.launch(share=True)
gradio_ui.launch(share=True)