app.py

import torch
import torchaudio
import gradio as gr
import os
import numpy as np

from transformers import Wav2Vec2Processor, Wav2Vec2Model
from safetensors.torch import load_file
import torch.nn as nn

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="creativepurus/accent-wav2vec2", filename="model.safetensors")

# Load processor
processor = Wav2Vec2Processor.from_pretrained("creativepurus/accent-wav2vec2")

# Load model weights from model.safetensors
state_dict = load_file(model_path, device="cpu")

# Define the same model architecture used during training
class Wav2Vec2Classifier(nn.Module):
    def __init__(self):
        super(Wav2Vec2Classifier, self).__init__()
        self.wav2vec2 = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-large-960h")
        self.dropout = nn.Dropout(0.3)
        self.classifier = nn.Linear(self.wav2vec2.config.hidden_size, 2)

    def forward(self, input_values, attention_mask=None):
        outputs = self.wav2vec2(input_values, attention_mask=attention_mask)
        hidden_states = outputs.last_hidden_state
        pooled = hidden_states.mean(dim=1)
        pooled = self.dropout(pooled)
        logits = self.classifier(pooled)
        return logits

# Instantiate and load the model
model = Wav2Vec2Classifier()
model.load_state_dict(state_dict)
model.eval()

# Prediction function
def predict_accent(audio):
    waveform, sample_rate = torchaudio.load(audio)
    if sample_rate != 16000:
        resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
        waveform = resampler(waveform)
    input_values = processor(waveform.squeeze().numpy(), return_tensors="pt", sampling_rate=16000).input_values
    with torch.no_grad():
        logits = model(input_values)
        predicted_class_id = logits.argmax().item()
    label_map = {0: "Canadian English", 1: "England English"}
    return label_map[predicted_class_id]

# # Gradio UI
# interface = gr.Interface(
#     fn=predict_accent,
#     inputs=gr.Audio(sources=["upload", "microphone"], type="filepath", label="Upload or Record Audio (WAV)"),
#     outputs=gr.Textbox(label="Predicted Accent"),
#     title="Accent Classification",
#     description="This app classifies English accents as either Canadian or England using a fine-tuned Wav2Vec2 model.",
#     allow_flagging="never"
# )


# Gradio UI with gr.Blocks

# Gradio UI with gr.Blocks and Custom Styling

custom_css = """
#predict-btn {
    background-color: orange !important;
    color: white !important;
    font-weight: bold;
}
#author-section {
    font-size: 18px;
    font-weight: 500;
}
"""

with gr.Blocks(css=custom_css) as demo:
    gr.Markdown("## 🗣️ Accent Classification App")
    gr.Markdown("This app classifies English accents as either **Canadian** or **England** using a fine-tuned Wav2Vec2 model.")
    
    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(sources=["upload", "microphone"], type="filepath", label="Upload or Record Audio (WAV)")
            predict_button = gr.Button("Predict Accent", elem_id="predict-btn")
        with gr.Column():
            result_output = gr.Textbox(label="Predicted Accent")

    predict_button.click(fn=predict_accent, inputs=audio_input, outputs=result_output)

    gr.Markdown("---")
    gr.Markdown("""
    <div id="author-section">
    👨‍💻 Created by <strong>Anand Purushottam</strong>  
    🔗 <a href="https://github.com/creativepurus" target="_blank">GitHub</a> | 
    <a href="https://linkedin.com/in/creativepurus" target="_blank">LinkedIn</a>
    </div>
    """,
    )

demo.launch()