app.py

import gradio as gr
from huggingface_hub import hf_hub_download
from audio_index import AudioEmbeddingSystem
from search import search, get_prompt
import pandas as pd
import numpy as np

db_file = hf_hub_download(
    repo_id="freddyaboulton/common-voice-english-audio",
    filename="audio_db_full.sqlite",
    repo_type="dataset",
)
index_file = hf_hub_download(
    repo_id="freddyaboulton/common-voice-english-audio",
    filename="audio_faiss_full.index",
    repo_type="dataset",
)

audio_embedding_system = AudioEmbeddingSystem(db_path=db_file, index_path=index_file)


def audio_search(audio_tuple, prompt: str):
    if audio_tuple is None:
        return gr.skip()

    sample_rate, array = audio_tuple
    if array.dtype == np.int16:
        array = array.astype(np.float32) / 32768.0

    rows = audio_embedding_system.search((sample_rate, array))
    print(rows)
    orig_rows = search(rows)
    for row in rows:
        path = row["path"]
        for orig in orig_rows:
            orig_row = orig["row"]
            print(orig_row)
            if orig_row["path"] == path:
                row["sentence"] = orig_row["sentence"]
                row["audio"] = [
                    "<audio src=" + orig_row["audio"][0]["src"] + " controls />"
                ]
    return pd.DataFrame(rows)[["path", "audio", "sentence", "distance"]].sort_values(
        by="distance", ascending=True
    )

sample_text = gr.Textbox(
    label="Prompt",
    info="Hit Enter to get a prompt from the common voice dataset",
    value=get_prompt(),
)
iface = gr.Interface(
    fn=audio_search,
    inputs=[gr.Audio(
        label="Record or upload a clip of your voice", sources=["microphone", "upload"]
    ), sample_text],
    outputs=gr.Dataframe(
        show_label=False,
        headers=["path", "audio", "sentence", "distance"],
        datatype=["str", "html", "str", "number"],
    ),
)
with gr.Blocks() as demo:
    gr.HTML(
        f"""
        <h1 style='text-align: center; display: flex; align-items: center; justify-content: center;'>
        <img src="/gradio_api/file=Karaoke_Huggy.png" alt="Voice Match" style="height: 100px; margin-right: 10px"> Voice Match
        </h1>
        """
    )
    gr.HTML(
    """
    <h2 style='text-align: center'>
    Powered by <a href="https://huggingface.co/rimelabs/rimecaster">RimeCaster</a>
    </h2>
    """
    )
    gr.Markdown(
        f"""
        <div style='text-align: center'>
        Record or upload an English clip of your voice and we'll find the most similar voices in the <a href="https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0">Common Voice dataset</a>.
        </div>
        """
    )
    iface.render()
    sample_text.submit(fn=get_prompt, inputs=None, outputs=sample_text)

demo.launch(allowed_paths=["Karaoke_Huggy.png"])