app.py

import gradio as gr
import os
from pathlib import Path
from vogent_turn.inference import TurnDetector
import soundfile as sf
import numpy as np


def get_detector():
    """Lazy load the detector to avoid initialization during import"""
    detector = TurnDetector(compile_model=False, warmup=False)
    return detector

# Initialize the turn detector
detector = get_detector()

# Get all preset names from samples folder
def get_presets():
    samples_dir = Path("samples")
    if not samples_dir.exists():
        return []
    presets = [d.name for d in samples_dir.iterdir() if d.is_dir()]
    return sorted(presets)

# Load preset data
def load_preset(preset_name):
    """Load audio and text files from the selected preset"""
    if not preset_name:
        return None, "", ""

    preset_dir = Path("samples") / preset_name

    # Load audio
    audio_path = preset_dir / "audio.wav"
    audio_file = str(audio_path) if audio_path.exists() else None

    # Load text files
    prev_text = ""
    curr_text = ""

    prev_path = preset_dir / "prev.txt"
    if prev_path.exists():
        prev_text = prev_path.read_text().strip()

    text_path = preset_dir / "text.txt"
    if text_path.exists():
        curr_text = text_path.read_text().strip()

    return audio_file, prev_text, curr_text

# Run inference
def run_inference(audio_file, prev_text, curr_text):
    """Run turn detection inference"""
    if audio_file is None:
        return "Error: No audio file provided"

    if curr_text is None:
        return "Error: a transcript of the audio must be provided"

    if prev_text is None:
        prev_text = ""

    try:
        # Load audio file
        audio, sr = sf.read(audio_file)

        # Convert to mono if stereo
        if len(audio.shape) > 1:
            audio = audio.mean(axis=1)

        # Ensure audio is float32
        audio = audio.astype(np.float32)

        # Run prediction with context
        result = detector.predict(
            audio,
            prev_line=prev_text if prev_text else None,
            curr_line=curr_text if curr_text else None,
            return_probs=True,
            sample_rate=sr,
        )

        # Format output
        is_endpoint = result['is_endpoint']
        prob_endpoint = result['prob_endpoint']
        prob_continue = result['prob_continue']

        output = f"""
**Turn Detection Result:**

- **Prediction:** {'Turn Complete (Endpoint)' if is_endpoint else 'Turn Incomplete (Continue)'}
- **Probability of Endpoint:** {prob_endpoint:.4f}
- **Probability of Continue:** {prob_continue:.4f}
"""
        return output

    except Exception as e:
        return f"Error during inference: {str(e)}"

# Get default preset and load its data
presets = get_presets()
default_preset = presets[0] if presets else None
default_audio, default_prev_text, default_curr_text = load_preset(default_preset) if default_preset else (None, "", "")

# Create Gradio interface
with gr.Blocks(title="Vogent Turn Demo") as demo:
    gr.Markdown("# Vogent Turn Demo")
    gr.Markdown("Multimodal turn detection using audio and text context")
    gr.Markdown("""
[GitHub](https://github.com/vogent/vogent-turn) | [Technical Report](https://blog.vogent.ai/posts/voturn-80m-state-of-the-art-turn-detection-for-voice-agents) | [Model Weights](http://huggingface.co/vogent/Vogent-Turn-80M)
""")

    with gr.Row():
        with gr.Column():
            # Preset selector
            preset_dropdown = gr.Dropdown(
                choices=presets,
                label="Preset Samples",
                info="Select a preset to auto-fill the fields",
                value=default_preset
            )

            # Input fields
            prev_text_input = gr.Textbox(
                label="Previous Line (The previous line spoken in the dialog)",
                placeholder="Enter the previous line of dialog...",
                lines=2,
                value=default_prev_text
            )

            curr_text_input = gr.Textbox(
                label="Current Line (The transcript of the below audio, omit punctuation)",
                placeholder="Enter the current line being spoken...",
                lines=2,
                value=default_curr_text
            )

            audio_input = gr.Audio(
                label="Audio",
                type="filepath",
                value=default_audio
            )

            # Inference button
            inference_btn = gr.Button("Run Inference", variant="primary")

        with gr.Column():
            # Output
            output_text = gr.Markdown(label="Results")

    # Connect preset dropdown to load function
    preset_dropdown.change(
        fn=load_preset,
        inputs=[preset_dropdown],
        outputs=[audio_input, prev_text_input, curr_text_input]
    )

    # Connect inference button
    inference_btn.click(
        fn=run_inference,
        inputs=[audio_input, prev_text_input, curr_text_input],
        outputs=[output_text]
    )

if __name__ == "__main__":
    demo.launch()