app.py

import os
import sys
import subprocess
import tempfile
import requests
from moviepy.editor import VideoFileClip

try:
    import whisper
    if not hasattr(whisper, 'load_model'):
        raise ImportError
except ImportError:
    subprocess.run([sys.executable, "-m", "pip", "install", "--upgrade", "openai-whisper"], check=True)
    import whisper

import torch
import librosa
import pandas as pd
from transformers import Wav2Vec2Processor, Wav2Vec2ForSequenceClassification
from huggingface_hub import login
import gradio as gr


# device = 'cuda' if torch.cuda.is_available() else 'cpu'
device = 'cpu'

def load_models():
    whisper_model = whisper.load_model('tiny', device=device)
    processor = Wav2Vec2Processor.from_pretrained(
        'jonatasgrosman/wav2vec2-large-english'
    )
    accent_model = Wav2Vec2ForSequenceClassification.from_pretrained(
        'jonatasgrosman/wav2vec2-large-english'
    ).to(device)
    accent_model = torch.quantization.quantize_dynamic(
          accent_model, {torch.nn.Linear}, dtype=torch.qint8
    )
    return whisper_model, processor, accent_model

whisper_model, processor, accent_model = load_models()

def analyze(video_url: str):
    with tempfile.NamedTemporaryFile(delete=False, suffix='.mp4') as tmp_vid:
        response = requests.get(video_url, stream=True)
        response.raise_for_status()
        for chunk in response.iter_content(chunk_size=1024 * 1024):
            if chunk:
                tmp_vid.write(chunk)
        video_path = tmp_vid.name

    audio_path = video_path.replace('.mp4', '.wav')
    clip = VideoFileClip(video_path)
    clip.audio.write_audiofile(audio_path, verbose=False, logger=None)
    clip.close()

    speech, sr = librosa.load(audio_path, sr=16000)

    result = whisper_model.transcribe(speech)
    transcript = result.get('text', '')
    lang = result.get('language', 'unknown')
    if lang != 'en':
        transcript = f"[Non-English detected: {lang}]\n" + transcript

    inputs = processor(speech, sampling_rate=sr, return_tensors='pt', padding=True)
    input_values = inputs.input_values.to(device)
    attention_mask = inputs.attention_mask.to(device)
    with torch.no_grad():
        logits = accent_model(input_values=input_values, attention_mask=attention_mask).logits
        probs = torch.softmax(logits, dim=-1).squeeze().cpu().tolist()

    accent_labels = [
        'American', 'Australian', 'British', 'Canadian', 'Indian',
        'Irish', 'New Zealander', 'South African', 'Welsh'
    ]  
    accent_probs = [(accent_labels[i], probs[i] * 100) for i in range(len(probs))]
    accent_probs.sort(key=lambda x: x[1], reverse=True)
    top_accent, top_conf = accent_probs[0]

    df = pd.DataFrame(accent_probs, columns=['Accent', 'Confidence (%)'])
    df = pd.DataFrame(accent_probs, columns=['Accent', 'Confidence (%)'])

    try:
        os.remove(video_path)
        os.remove(audio_path)
    except:
        pass

    return top_accent, f"{top_conf:.2f}%", df

interface = gr.Interface(
    fn=analyze,
    inputs=gr.Textbox(label='Video URL', placeholder='Enter public MP4 URL'),
    outputs=[
        # gr.Textbox(label='Transcript'),
        gr.Textbox(label='Predicted Accent'),
        gr.Textbox(label='Accent Confidence'),
        gr.Dataframe(label='All Accent Probabilities')
    ],
    title='English Accent Detector',
    description='Paste a direct MP4 URL to extract, transcribe, and classify English accents. It is a bit slow since we run Whisper and Wav2Vec2 models on CPU. Please test with short videos.',
    examples=[
        ['http://commondatastorage.googleapis.com/gtv-videos-bucket/sample/ForBiggerBlazes.mp4'],
    ],
    allow_flagging='never'
)

if __name__ == '__main__':
    interface.launch()