app.py

import os
import torchaudio
import gradio as gr
import spaces
import torch
from transformers import AutoProcessor, AutoModelForCTC

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")

# load examples 
examples = []
examples_dir = "examples"
if os.path.exists(examples_dir):
    for filename in os.listdir(examples_dir):
        if filename.endswith((".wav", ".mp3", ".ogg")):
            examples.append([os.path.join(examples_dir, filename)])

# Load model and processor
MODEL_PATH = "badrex/w2v-bert-2.0-kinyarwanda-asr"
processor = AutoProcessor.from_pretrained(MODEL_PATH)
model = AutoModelForCTC.from_pretrained(MODEL_PATH)

# move model and processor to device
model = model.to(device)

@spaces.GPU()
def process_audio(audio_path):
    """Process audio with return the generated response.
    
    Args:
        audio_path: Path to the audio file to be transcribed.    
    Returns:
        String containing the transcribed text from the audio file, or an error message
        if the audio file is missing.
    """
    if not audio_path:
        return "Please upload an audio file."

    # get audio array
    audio_array, sample_rate = torchaudio.load(audio_path)

    # if sample rate is not 16000, resample to 16000
    if sample_rate != 16000:
        audio_array = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)(audio_array)

    inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        logits = model(**inputs).logits

    outputs = torch.argmax(logits, dim=-1)
    
    decoded_outputs = processor.batch_decode(
        outputs,
        skip_special_tokens=True
    )
    
    return decoded_outputs[0].strip()


# Define Gradio interface
with gr.Blocks(title="<div>ASRwanda 🎙️ <br>Speech Recognition for Kinyarwanda</div>") as demo:
    gr.Markdown("""
        <div class="centered-content">
            <div>
                <p>
                Developed with ❤ by <a href="https://badrex.github.io/" style="color: #2563eb;">Badr al-Absi</a> ☕
                </p>
                <br>
                <p style="font-size: 15px; line-height: 1.8;">
                 Muraho 👋🏼
                <br><br>
                 This is a demo for ASRwanda, a Transformer-based automatic speech recognition (ASR) system for Kinyarwanda language.
                 The underlying ASR model was trained on 1000 hours of transcribed speech provided by 
                 <a href="https://digitalumuganda.com/" style="color: #2563eb;">Digital Umuganda</a> as part of the Kinyarwanda
                 <a href="https://www.kaggle.com/competitions/kinyarwanda-automatic-speech-recognition-track-b" style="color: #2563eb;"> ASR hackathon</a> on Kaggle.
                <br><br>                   
                Simply <strong>upload an audio file</strong> 📤 or <strong>record yourself speaking</strong> 🎙️⏺️ to try out the model!
                </p>
            </div>
        </div>
    """)

    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(type="filepath", label="Upload Audio")
            submit_btn = gr.Button("Transcribe Audio", variant="primary")
        
        with gr.Column():
            output_text = gr.Textbox(label="Text Transcription", lines=10)
    
    submit_btn.click(
        fn=process_audio,
        inputs=[audio_input],
        outputs=output_text
    )

    gr.Examples(
        examples=examples if examples else None,
        inputs=[audio_input],
    )

# Launch the app
if __name__ == "__main__":
    demo.queue().launch()