app.py

import gradio as gr
import torch
import librosa
from transformers import Wav2Vec2Processor, AutoModelForCTC
import zipfile
import os
import firebase_admin
from firebase_admin import credentials, firestore
from datetime import datetime
import json
import tempfile

# Initialize Firebase
firebase_config = json.loads(os.environ.get('firebase_creds'))
cred = credentials.Certificate(firebase_config)  # Your Firebase JSON key file
firebase_admin.initialize_app(cred)
db = firestore.client()

# Load the ASR model and processor
MODEL_NAME = "eleferrand/xlsr53_Amis"
processor = Wav2Vec2Processor.from_pretrained(MODEL_NAME)
model = AutoModelForCTC.from_pretrained(MODEL_NAME)

def transcribe(audio_file):
    """
    Transcribes the audio file using the loaded ASR model.
    Returns the transcription string.
    """
    try:
        # Load and resample the audio to 16kHz
        audio, rate = librosa.load(audio_file, sr=16000)
        input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_values

        # Get model predictions and decode to text
        with torch.no_grad():
            logits = model(input_values).logits
        predicted_ids = torch.argmax(logits, dim=-1)
        transcription = processor.batch_decode(predicted_ids)[0]
        return transcription.replace("[UNK]", "")
    except Exception as e:
        return f"Error processing file: {e}"

def transcribe_both(audio_file):
    """
    Transcribes the audio and returns:
      - the original transcription (non-editable textbox),
      - the transcription (pre-filled for the editable textbox), and
      - the processing time (in seconds).
    """
    start_time = datetime.now()
    transcription = transcribe(audio_file)
    processing_time = (datetime.now() - start_time).total_seconds()
    return transcription, transcription, processing_time

def store_correction(original_transcription, corrected_transcription, audio_file, processing_time, age, native_speaker):
    """
    Stores the transcriptions and additional metadata (including user info and audio details)
    in Firestore as a single document.
    """
    try:
        audio_metadata = {}
        if audio_file and os.path.exists(audio_file):
            audio, sr = librosa.load(audio_file, sr=16000)
            duration = librosa.get_duration(y=audio, sr=sr)
            file_size = os.path.getsize(audio_file)
            audio_metadata = {'duration': duration, 'file_size': file_size}
        combined_data = {
            'original_text': original_transcription,
            'corrected_text': corrected_transcription,
            'timestamp': datetime.now().isoformat(),
            'processing_time': processing_time,
            'audio_metadata': audio_metadata,
            'audio_url': None,  # Placeholder if you decide to store an URL later
            'model_name': MODEL_NAME,
            'user_info': {
                'native_amis_speaker': native_speaker,
                'age': age
            }
        }
        db.collection('transcriptions').add(combined_data)
        return "Correction saved successfully!"
    except Exception as e:
        return f"Error saving correction: {e}"

def prepare_download(audio_file, original_transcription, corrected_transcription):
    """
    Prepares a ZIP file containing:
      - the uploaded audio file (as audio.wav),
      - a text file with the original transcription, and
      - a text file with the corrected transcription.
    Returns the ZIP file's path.
    """
    if audio_file is None:
        return None

    # Create a temporary file to avoid filename conflicts
    tmp_zip = tempfile.NamedTemporaryFile(delete=False, suffix=".zip")
    tmp_zip.close()
    with zipfile.ZipFile(tmp_zip.name, "w") as zf:
        # Add the audio file (renamed inside the zip)
        if os.path.exists(audio_file):
            zf.write(audio_file, arcname="audio.wav")
        else:
            print("Audio file not found:", audio_file)

        # Write and add the original transcription
        orig_txt = "original_transcription.txt"
        with open(orig_txt, "w", encoding="utf-8") as f:
            f.write(original_transcription)
        zf.write(orig_txt, arcname="original_transcription.txt")
        os.remove(orig_txt)

        # Write and add the corrected transcription
        corr_txt = "corrected_transcription.txt"
        with open(corr_txt, "w", encoding="utf-8") as f:
            f.write(corrected_transcription)
        zf.write(corr_txt, arcname="corrected_transcription.txt")
        os.remove(corr_txt)
    return tmp_zip.name

# Build the Gradio Blocks interface with improved styling
with gr.Blocks(css="""
    .container { 
        max-width: 800px; 
        margin: auto; 
        padding: 20px; 
        font-family: Arial, sans-serif;
    }
    .header { 
        text-align: center; 
        margin-bottom: 30px; 
    }
    .section { 
        margin-bottom: 30px; 
        padding: 15px; 
        border: 1px solid #ddd; 
        border-radius: 8px; 
        background-color: #f9f9f9;
    }
    .section h3 {
        margin-top: 0;
        margin-bottom: 15px;
        text-align: center;
    }
    .button-row {
        display: flex;
        justify-content: center;
        gap: 10px;
        flex-wrap: wrap;
    }
    @media (max-width: 600px) {
      .gradio-row { 
          flex-direction: column; 
      }
    }
""") as demo:
    with gr.Column(elem_classes="container"):
        gr.Markdown("<h1 class='header'>ASR Demo with Editable Transcription</h1>")
        
        # Step 1: Audio Upload & Transcription
        with gr.Box(elem_classes="section"):
            gr.Markdown("### Step 1: Audio Upload & Transcription")
            with gr.Row(elem_classes="gradio-row"):
                audio_input = gr.Audio(sources=["upload", "microphone"], type="filepath", label="Audio Input")
                transcribe_button = gr.Button("Transcribe Audio", variant="primary")
            proc_time_state = gr.State()
        
        # Step 2: Review & Edit Transcription
        with gr.Box(elem_classes="section"):
            gr.Markdown("### Step 2: Review & Edit Transcription")
            with gr.Row(elem_classes="gradio-row"):
                original_text = gr.Textbox(label="Original Transcription", interactive=False, lines=5, placeholder="Transcription will appear here...")
                corrected_text = gr.Textbox(label="Corrected Transcription", interactive=True, lines=5, placeholder="Edit transcription here...")
        
        # Step 3: User Information
        with gr.Box(elem_classes="section"):
            gr.Markdown("### Step 3: User Information")
            with gr.Row(elem_classes="gradio-row"):
                age_input = gr.Slider(minimum=0, maximum=100, step=1, label="Age", value=25)
                native_speaker_input = gr.Checkbox(label="Native Amis Speaker", value=True)
        
        # Step 4: Save & Download
        with gr.Box(elem_classes="section"):
            gr.Markdown("### Step 4: Save & Download")
            with gr.Row(elem_classes="button-row"):
                save_button = gr.Button("Save Correction to Database", variant="primary")
                save_status = gr.Textbox(label="Save Status", interactive=False, placeholder="Status messages will appear here...")
            with gr.Row(elem_classes="button-row"):
                download_button = gr.Button("Download Results (ZIP)")
                download_output = gr.File(label="Download ZIP")
        
        # UI Actions
        transcribe_button.click(
            fn=transcribe_both,
            inputs=audio_input,
            outputs=[original_text, corrected_text, proc_time_state]
        )
        
        save_button.click(
            fn=store_correction,
            inputs=[original_text, corrected_text, audio_input, proc_time_state, age_input, native_speaker_input],
            outputs=save_status
        )
        
        download_button.click(
            fn=prepare_download,
            inputs=[audio_input, original_text, corrected_text],
            outputs=download_output
        )


# Launch the demo
demo.launch(share=True)