app.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

# This file is copied & modified from UNESCO/MMS/blob/main/asr.py

import gradio as gr
import numpy as np
import torch
import torchaudio
import json
from audiobox_aesthetics import infer as aes_infer
import plotly.graph_objects as go

ASR_EXAMPLES = [("assets/bach.wav")]

aes_predictor = aes_infer.initialize_model(None)


def transcribe(audio_data=None):
    if not audio_data:
        return "<<ERROR: Empty Audio Input>>"

    if isinstance(audio_data, tuple):
        # microphone
        sr, audio_samples = audio_data
        audio_samples = (audio_samples / 32768.0).astype(np.float32)
        audio_samples = torch.tensor(audio_samples)
        if audio_samples.ndim == 1:
            audio_samples = audio_samples[:, None]

        assert audio_samples.ndim == 2
        audio_samples = audio_samples.t()
    else:
        # file upload
        if not isinstance(audio_data, str):
            return "<<ERROR: Invalid Audio Input Instance: {}>>".format(
                type(audio_data)
            )
        audio_samples, sr = torchaudio.load(audio_data)

    transcription = json.loads(
        aes_predictor.forward([{"path": audio_samples, "sample_rate": sr}])[0]
    )

    # reorder like paper figures
    transcription = {
        "Production Quality": transcription["PQ"],
        "Production Complexity": transcription["PC"],
        "Content Enjoyment": transcription["CE"],
        "Content Usefulness": transcription["CU"],
    }

    # Create a Plotly bar plot
    fig = go.Figure()
    colors = ["#b1d8ff", "#fee2f5", "#cefac4", "#d2d3ff"]
    values = list(transcription.values())
    keys = list(transcription.keys())

    fig.add_trace(
        go.Bar(
            x=keys,
            y=values,
            text=[f"{v:.2f}" for v in values],  # Format text to 2 decimal places
            textposition="outside",  # Position text outside the bars
            marker=dict(color=colors),
        )
    )

    # Set the range for the y-axis
    fig.update_layout(
        yaxis=dict(range=[0, 10]),
        xaxis_title="Metrics",
        yaxis_title="Scores",
    )
    return fig


main_interface = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(),
    ],
    # outputs="text",
    outputs=gr.Plot(),
    examples=ASR_EXAMPLES,
    title="Audiobox Aesthetics Demo Prediction",
    description=("Play some audio through microphone or upload the file."),
    article="",
    allow_flagging="never",
)

disclaimer = """
## Disclaimer
"""

with gr.Blocks() as demo:
    gr.HTML(
        """
            <div style="text-align: center;">
            <h1>
                Meta Audiobox Aesthetics: Unified Automatic Quality Assessment for Speech, Music, and Sound
            </h1>
            </div>
           """
    )

    gr.Markdown(
        "<p align='center' style='font-size: 20px;'>See our <a href='https://arxiv.org/abs/2502.05139'>paper</a>, Github <a href='https://github.com/facebookresearch/audiobox-aesthetics'>repo</a> and HuggingFace <a href='https://huggingface.co/facebook/audiobox-aesthetics'>repo</a> </p>"
    )
    gr.HTML(
        """<center><a href="https://huggingface.co/spaces/facebook/audiobox-aesthetics?duplicate=true"  style="display: inline-block;margin-top: .5em;margin-right: .25em;" target="_blank"><img style="margin-bottom: 0em;display: inline;margin-top: -.25em;" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> for more control and no queue.</center>"""
    )

    main_interface.render()

    gr.HTML(
        """
            <div class="footer" style="text-align:center">
                <p>
                    Model by <a href="https://ai.facebook.com" style="text-decoration: underline;" target="_blank">Meta AI</a> - Gradio Demo by 🤗 Hugging Face
                </p>
            </div>
           """
    )

    # with gr.Row():
    #     gr.Markdown(disclaimer)

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