app.py

import gradio as gr
import torch
from diffusers import AudioLDM2Pipeline


# make Space compatible with CPU duplicates
if torch.cuda.is_available():
    device = "cuda"
    torch_dtype = torch.float16
else:
    device = "cpu"
    torch_dtype = torch.float32

# load the diffusers pipeline
repo_id = "cvssp/audioldm2"
pipe = AudioLDM2Pipeline.from_pretrained(repo_id, torch_dtype=torch_dtype).to(device)
# pipe.unet = torch.compile(pipe.unet)

# set the generator for reproducibility
generator = torch.Generator(device)


def text2audio(text, negative_prompt, duration, guidance_scale, random_seed, n_candidates):
    if text is None:
        raise gr.Error("Please provide a text input.")

    waveforms = pipe(
        text,
        audio_length_in_s=duration,
        guidance_scale=guidance_scale,
        num_inference_steps=200,
        negative_prompt=negative_prompt,
        num_waveforms_per_prompt=n_candidates if n_candidates else 1,
        generator=generator.manual_seed(int(random_seed)),
    )["audios"]

    return gr.make_waveform((16000, waveforms[0]), bg_image="bg.png")


iface = gr.Blocks()

with iface:
    gr.HTML(
        """
            <div style="text-align: center; max-width: 700px; margin: 0 auto;">
              <div
                style="
                  display: inline-flex; align-items: center; gap: 0.8rem; font-size: 1.75rem;
                "
              >
                <h1 style="font-weight: 900; margin-bottom: 7px; line-height: normal;">
                  AudioLDM 2: A General Framework for Audio, Music, and Speech Generation
                </h1>
              </div> <p style="margin-bottom: 10px; font-size: 94%">
                <a href="https://arxiv.org/abs/2308.05734">[Paper]</a> <a href="https://audioldm.github.io/audioldm2">[Project
                page]</a> <a href="https://huggingface.co/docs/diffusers/main/en/api/pipelines/audioldm2">[🧨
                Diffusers]</a>
              </p>
            </div>
        """
    )
    gr.HTML("""This is the demo for AudioLDM 2, powered by 🧨 Diffusers. Demo uses the checkpoint <a
        href="https://huggingface.co/cvssp/audioldm2"> AudioLDM 2 base</a>. For faster inference without waiting in
        queue, you may duplicate the space and upgrade to a GPU in the settings.""")
    gr.DuplicateButton()

    with gr.Group():
        textbox = gr.Textbox(
            value="The vibrant beat of Brazilian samba drums.",
            max_lines=1,
            label="Input text",
            info="Your text is important for the audio quality. Please ensure it is descriptive by using more adjectives.",
            elem_id="prompt-in",
        )
        negative_textbox = gr.Textbox(
            value="Low quality.",
            max_lines=1,
            label="Negative prompt",
            info="Enter a negative prompt not to guide the audio generation. Selecting appropriate negative prompts can improve the audio quality significantly.",
            elem_id="prompt-in",
        )

        with gr.Accordion("Click to modify detailed configurations", open=False):
            seed = gr.Number(
                value=45,
                label="Seed",
                info="Change this value (any integer number) will lead to a different generation result.",
            )
            duration = gr.Slider(5, 15, value=10, step=2.5, label="Duration (seconds)")
            guidance_scale = gr.Slider(
                0,
                7,
                value=3.5,
                step=0.5,
                label="Guidance scale",
                info="Larger => better quality and relevancy to text; Smaller => better diversity",
            )
            n_candidates = gr.Slider(
                1,
                5,
                value=3,
                step=1,
                label="Number waveforms to generate",
                info="Automatic quality control. This number control the number of candidates (e.g., generate three audios and choose the best to show you). A larger value usually lead to better quality with heavier computation",
            )

        outputs = gr.Video(label="Output", elem_id="output-video")
        btn = gr.Button("Submit")

    btn.click(
        text2audio,
        inputs=[textbox, negative_textbox, duration, guidance_scale, seed, n_candidates],
        outputs=[outputs],
    )

    gr.HTML(
        """
    <div class="footer" style="text-align: center">
        <p>Share your generations with the community by clicking the share icon at the top right the generated audio!</p>
        <p>Follow the latest update of AudioLDM 2 on our<a href="https://audioldm.github.io/audioldm2"
        style="text-decoration: underline;" target="_blank"> Github repo</a> </p> 
        <p>Model by <a
        href="https://twitter.com/LiuHaohe" style="text-decoration: underline;" target="_blank">Haohe
        Liu</a>. Code and demo by 🤗 Hugging Face.</p>
    </div>
    """
    )
    gr.Examples(
        [
            ["A hammer is hitting a wooden surface.", "Low quality.", 10, 3.5, 45, 3],
            ["A cat is meowing for attention.", "Low quality.", 10, 3.5, 45, 3],
            ["An excited crowd cheering at a sports game.", "Low quality.", 10, 3.5, 45, 3],
            ["Birds singing sweetly in a blooming garden.", "Low quality.", 10, 3.5, 45, 3],
            ["A modern synthesizer creating futuristic soundscapes.", "Low quality.", 10, 3.5, 45, 3],
            ["The vibrant beat of Brazilian samba drums.", "Low quality.", 10, 3.5, 45, 3],
        ],
        fn=text2audio,
        inputs=[textbox, negative_textbox, duration, guidance_scale, seed, n_candidates],
        outputs=[outputs],
        cache_examples=True,
    )
    gr.HTML(
        """
            <div class="acknowledgements"> <p>Essential Tricks for Enhancing the Quality of Your Generated
            Audio</p>
            <p>1. Try using more adjectives to describe your sound. For example: "A man is speaking
            clearly and slowly in a large room" is better than "A man is speaking".</p>
            <p>2. Try using different random seeds, which can significantly affect the quality of the generated 
            output.</p>
            <p>3. It's better to use general terms like 'man' or 'woman' instead of specific names for individuals or 
            abstract objects that humans may not be familiar with.</p>
            <p>4. Using a negative prompt to not guide the diffusion process can improve the
            audio quality significantly. Try using negative prompts like 'low quality'.</p>
            </div>
            """
    )
    with gr.Accordion("Additional information", open=False):
        gr.HTML(
            """
            <div class="acknowledgments">
                <p> We build the model with data from <a href="http://research.google.com/audioset/">AudioSet</a>,
                <a href="https://freesound.org/">Freesound</a> and <a
                href="https://sound-effects.bbcrewind.co.uk/">BBC Sound Effect library</a>. We share this demo
                based on the <a
                href="https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/375954/Research.pdf">UK
                copyright exception</a> of data for academic research. 
                </p>
            </div>
            """
        )

iface.queue(max_size=20).launch()