app.py

import gradio as gr
import torch
from nemo.collections.asr.models import EncDecSpeakerLabelModel


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

STYLE = """
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" integrity="sha256-YvdLHPgkqJ8DVUxjjnGVlMMJtNimJ6dYkowFFvp4kKs=" crossorigin="anonymous">
"""
OUTPUT_OK = (
    STYLE
    + """
    <div class="container">
        <div class="row"><h1 style="text-align: center">The provided samples are</h1></div>
        <div class="row"><h1 class="text-success" style="text-align: center">Same Speakers!!!</h1></div>
    </div>
"""
)
OUTPUT_FAIL = (
    STYLE
    + """
    <div class="container">
        <div class="row"><h1 style="text-align: center">The provided samples are from </h1></div>
        <div class="row"><h1 class="text-danger" style="text-align: center">Different Speakers!!!</h1></div>       
    </div>
"""
)

THRESHOLD = 0.80

model_name = "nvidia/speakerverification_en_titanet_large"
model = EncDecSpeakerLabelModel.from_pretrained(model_name).to(device)


def compare_samples(path1, path2):
    if not (path1 and path2):
        return '<b style="color:red">ERROR: Please record audio for *both* speakers!</b>'

    output = model.verify_speakers(path1,path2,THRESHOLD)

    return OUTPUT_OK if output else OUTPUT_FAIL


inputs = [
    gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #1"),
    gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #2"),
]
output = gr.outputs.HTML(label="")


description = (
    "This demonstration will analyze two recordings of speech and ascertain whether they have been spoken by the same individual.\n"
    "You can attempt this exercise using your own voice."
)
article = (
    "<p style='text-align: center'>"
    "<a href='https://huggingface.co/nvidia/speakerverification_en_titanet_large' target='_blank'>🎙️ Learn more about TitaNet model</a> | "
    "<a href='https://arxiv.org/pdf/2110.04410.pdf' target='_blank'>📚 TitaNet paper</a> | "
    "<a href='https://github.com/NVIDIA/NeMo' target='_blank'>🧑‍💻 Repository</a>"
    "</p>"
)
examples = [
    ["data/id10270_5r0dWxy17C8-00001.wav", "data/id10270_5r0dWxy17C8-00002.wav"],
    ["data/id10271_1gtz-CUIygI-00001.wav", "data/id10271_1gtz-CUIygI-00002.wav"],
    ["data/id10270_5r0dWxy17C8-00001.wav", "data/id10271_1gtz-CUIygI-00001.wav"],
    ["data/id10270_5r0dWxy17C8-00002.wav", "data/id10271_1gtz-CUIygI-00002.wav"],
]

interface = gr.Interface(
    fn=compare_samples,
    inputs=inputs,
    outputs=output,
    title="Speaker Verification with TitaNet Embeddings",
    description=description,
    article=article,
    layout="horizontal",
    theme="huggingface",
    allow_flagging=False,
    live=False,
    examples=examples,
)
interface.launch(enable_queue=True)