app.py

import gradio as gr
import os
import traceback

# Load Hugging Face API token from environment variables
HUGGINGFACE_TOKEN = os.getenv("HF_TOKEN")

if not HUGGINGFACE_TOKEN:
    raise ValueError("Missing Hugging Face token. Please set HF_TOKEN in Space settings.")

# Function to transcribe audio
def transcribe_audio(audio):
    if audio is None:
        return "Error: Please upload or record an audio file."

    # Check file size
    file_size_mb = os.path.getsize(audio) / (1024 * 1024)
    if file_size_mb > 100:
        return f"Error: File size ({file_size_mb:.2f}MB) exceeds 100MB limit."

    try:
        global model
        if model is None:
            return "Error: Model failed to load."

        # Transcribe
        result = model(audio)
        print("Model output:", result)  # Debugging output

        # Handle output format
        if isinstance(result, dict) and 'text' in result:
            return result['text']
        elif isinstance(result, str):
            return result
        else:
            return f"Unexpected output from model: {result}"

    except Exception as e:
        error_str = str(e)
        if "3000 mel input features" in error_str:
            return "Error: Audio exceeds 30 seconds. Long-form transcription not supported in this configuration."
        return f"Error during transcription: {error_str}"

# Load model securely using HF_TOKEN
model = None
try:
    model = gr.load(
        "models/vinai/PhoWhisper-large",
        provider="hf-inference",
        api_key=HUGGINGFACE_TOKEN  # Pass token securely
    )
    print("Model loaded successfully!")
except Exception as e:
    print("Error loading model:", str(e))
    traceback.print_exc()

# Build Gradio UI
with gr.Blocks(fill_height=True) as demo:
    with gr.Sidebar():
        gr.Markdown("# Inference Provider")
        gr.Markdown("This Space showcases the vinai/PhoWhisper-large model, served by the hf-inference API.")
        button = gr.LoginButton("Sign in")

    with gr.Column():
        gr.Markdown("# Audio Transcription with PhoWhisper-large")

        audio_input = gr.Audio(
            sources=["upload", "microphone"],
            type="filepath",
            label="Upload an audio file or record (max 100MB)"
        )

        transcription_output = gr.Textbox(
            label="Transcription",
            lines=10,
            placeholder="Transcription or error message will appear here...",
            interactive=False
        )

        submit_btn = gr.Button("Transcribe")

        submit_btn.click(
            fn=transcribe_audio,
            inputs=audio_input,
            outputs=transcription_output,
            queue=True,
            concurrency_limit=2
        )

# Launch the app
demo.launch()