app.py

import gradio as gr
import collections
import scipy.signal
import numpy as np
import matplotlib.pyplot as plt
from functools import partial

from openwakeword.utils import download_models
from openwakeword.model import Model

# Download all ONNX + TFLite models once
download_models()

# Initialize the ONNX-based wake-word model
model = Model(inference_framework="onnx")

# Factory for per-model rolling buffers
initial_state = collections.defaultdict(partial(collections.deque, maxlen=60))

def process_audio(audio, state):
    sr, samples = audio
    # Resample if not 16 kHz
    if sr != 16000:
        samples = scipy.signal.resample(samples, int(len(samples) / sr * 16000))
    data = samples

    detected_msg = ""  # Will hold our “Detected X!” text

    # Slide in 1280-sample windows
    for i in range(0, len(data), 1280):
        chunk = data[i : i + 1280]
        # Stereo → mono
        if chunk.ndim == 2 and chunk.shape[1] == 2:
            chunk = chunk[:, 0]
        if len(chunk) == 1280:
            preds = model.predict(chunk)
            for name, score in preds.items():
                # Prime with zeros the first time
                if len(state[name]) == 0:
                    state[name].extend(np.zeros(60))
                state[name].append(score)

                # If you want a threshold trigger:
                if score > 0.8 and not detected_msg:
                    detected_msg = f"🗣 Detected **{name}**!"

    # Build the plot
    fig, ax = plt.subplots()
    for name, dq in state.items():
        ax.plot(np.arange(len(dq)), list(dq), label=name)

    ax.set_ylim(0, 1)
    ax.set_xlabel("Time (frames)")
    ax.set_ylabel("Model Score")

    # Only add a legend if at least one line has a label
    if state:
        ax.legend(loc="center left", bbox_to_anchor=(1, 0.5), fontsize="small")

    plt.tight_layout()

    # Capture and close to avoid memory leak
    out_fig = fig
    plt.close(fig)

    return out_fig, state, detected_msg

description = """
Speak one of the wake-words into your mic and watch its score spike!

| Model Name     | Phrase                             |
|----------------|------------------------------------|
| alexa          | "alexa"                            |
| hey_mycroft    | "hey mycroft"                      |
| hey_jarvis     | "hey jarvis"                       |
| hey_rhasspy    | "hey rhasspy"                      |
| weather        | "what's the weather"               |
| x_minute_timer | "set a timer for 1 minute"         |
"""

iface = gr.Interface(
    fn=process_audio,
    title="openWakeWord Live Demo [FIXED]",
    description=description,
    inputs=[
        gr.Audio(sources=["microphone"], type="numpy", streaming=True, show_label=False),
        gr.State(initial_state),
    ],
    outputs=[
        gr.Plot(label="Model Scores"),
        gr.State(),
        gr.Textbox(label="Detection", interactive=False),
    ],
    live=True,
    css=".flex {flex-direction: column} .gr-panel {width: 100%}"
)

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