{"cells": [{"cell_type": "markdown", "id": "302934307671667531413257853548643485645", "metadata": {}, "source": ["# Gradio Demo: same-person-or-different\n", "### This demo identifies if two speakers are the same person using Gradio's Audio and HTML components.\n", "        "]}, {"cell_type": "code", "execution_count": null, "id": "272996653310673477252411125948039410165", "metadata": {}, "outputs": [], "source": ["!pip install -q gradio git+https://github.com/huggingface/transformers torchaudio "]}, {"cell_type": "code", "execution_count": null, "id": "288918539441861185822528903084949547379", "metadata": {}, "outputs": [], "source": ["# Downloading files from the demo repo\n", "import os\n", "!wget -q https://github.com/gradio-app/gradio/raw/main/demo/same-person-or-different/packages.txt\n", "os.mkdir('samples')\n", "!wget -q -O samples/cate_blanch.mp3 https://github.com/gradio-app/gradio/raw/main/demo/same-person-or-different/samples/cate_blanch.mp3\n", "!wget -q -O samples/cate_blanch_2.mp3 https://github.com/gradio-app/gradio/raw/main/demo/same-person-or-different/samples/cate_blanch_2.mp3\n", "!wget -q -O samples/heath_ledger.mp3 https://github.com/gradio-app/gradio/raw/main/demo/same-person-or-different/samples/heath_ledger.mp3"]}, {"cell_type": "code", "execution_count": null, "id": "44380577570523278879349135829904343037", "metadata": {}, "outputs": [], "source": ["import gradio as gr\n", "import torch\n", "from torchaudio.sox_effects import apply_effects_file\n", "from transformers import AutoFeatureExtractor, AutoModelForAudioXVector\n", "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", "\n", "OUTPUT_OK = (\n", "    \"\"\"\n", "    <div class=\"container\">\n", "        <div class=\"row\"><h1 style=\"text-align: center\">The speakers are</h1></div>\n", "        <div class=\"row\"><h1 class=\"display-1 text-success\" style=\"text-align: center\">{:.1f}%</h1></div>\n", "        <div class=\"row\"><h1 style=\"text-align: center\">similar</h1></div>\n", "        <div class=\"row\"><h1 class=\"text-success\" style=\"text-align: center\">Welcome, human!</h1></div>\n", "        <div class=\"row\"><small style=\"text-align: center\">(You must get at least 85% to be considered the same person)</small><div class=\"row\">\n", "    </div>\n", "\"\"\"\n", ")\n", "OUTPUT_FAIL = (\n", "    \"\"\"\n", "    <div class=\"container\">\n", "        <div class=\"row\"><h1 style=\"text-align: center\">The speakers are</h1></div>\n", "        <div class=\"row\"><h1 class=\"display-1 text-danger\" style=\"text-align: center\">{:.1f}%</h1></div>\n", "        <div class=\"row\"><h1 style=\"text-align: center\">similar</h1></div>\n", "        <div class=\"row\"><h1 class=\"text-danger\" style=\"text-align: center\">You shall not pass!</h1></div>\n", "        <div class=\"row\"><small style=\"text-align: center\">(You must get at least 85% to be considered the same person)</small><div class=\"row\">\n", "    </div>\n", "\"\"\"\n", ")\n", "\n", "EFFECTS = [\n", "    [\"remix\", \"-\"],\n", "    [\"channels\", \"1\"],\n", "    [\"rate\", \"16000\"],\n", "    [\"gain\", \"-1.0\"],\n", "    [\"silence\", \"1\", \"0.1\", \"0.1%\", \"-1\", \"0.1\", \"0.1%\"],\n", "    [\"trim\", \"0\", \"10\"],\n", "]\n", "\n", "THRESHOLD = 0.85\n", "\n", "model_name = \"microsoft/unispeech-sat-base-plus-sv\"\n", "feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)\n", "model = AutoModelForAudioXVector.from_pretrained(model_name).to(device)\n", "cosine_sim = torch.nn.CosineSimilarity(dim=-1)\n", "\n", "\n", "def similarity_fn(path1, path2):\n", "    if not (path1 and path2):\n", "        return '<b style=\"color:red\">ERROR: Please record audio for *both* speakers!</b>'\n", "\n", "    wav1, _ = apply_effects_file(path1, EFFECTS)\n", "    wav2, _ = apply_effects_file(path2, EFFECTS)\n", "    print(wav1.shape, wav2.shape)\n", "\n", "    input1 = feature_extractor(wav1.squeeze(0), return_tensors=\"pt\", sampling_rate=16000).input_values.to(device)\n", "    input2 = feature_extractor(wav2.squeeze(0), return_tensors=\"pt\", sampling_rate=16000).input_values.to(device)\n", "\n", "    with torch.no_grad():\n", "        emb1 = model(input1).embeddings\n", "        emb2 = model(input2).embeddings\n", "    emb1 = torch.nn.functional.normalize(emb1, dim=-1).cpu()\n", "    emb2 = torch.nn.functional.normalize(emb2, dim=-1).cpu()\n", "    similarity = cosine_sim(emb1, emb2).numpy()[0]\n", "\n", "    if similarity >= THRESHOLD:\n", "        output = OUTPUT_OK.format(similarity * 100)\n", "    else:\n", "        output = OUTPUT_FAIL.format(similarity * 100)\n", "\n", "    return output\n", "\n", "inputs = [\n", "    gr.Audio(sources=[\"microphone\"], type=\"filepath\", label=\"Speaker #1\"),\n", "    gr.Audio(sources=[\"microphone\"], type=\"filepath\", label=\"Speaker #2\"),\n", "]\n", "output = gr.HTML(label=\"\")\n", "\n", "\n", "description = (\n", "    \"This demo from Microsoft will compare two speech samples and determine if they are from the same speaker. \"\n", "    \"Try it with your own voice!\"\n", ")\n", "article = (\n", "    \"<p style='text-align: center'>\"\n", "    \"<a href='https://huggingface.co/microsoft/unispeech-sat-large-sv' target='_blank'>\ud83c\udf99\ufe0f Learn more about UniSpeech-SAT</a> | \"\n", "    \"<a href='https://arxiv.org/abs/2110.05752' target='_blank'>\ud83d\udcda UniSpeech-SAT paper</a> | \"\n", "    \"<a href='https://www.danielpovey.com/files/2018_icassp_xvectors.pdf' target='_blank'>\ud83d\udcda X-Vector paper</a>\"\n", "    \"</p>\"\n", ")\n", "examples = [\n", "    [\"samples/cate_blanch.mp3\", \"samples/cate_blanch_2.mp3\"],\n", "    [\"samples/cate_blanch.mp3\", \"samples/heath_ledger.mp3\"],\n", "]\n", "\n", "interface = gr.Interface(\n", "    fn=similarity_fn,\n", "    inputs=inputs,\n", "    outputs=output,\n", "    layout=\"horizontal\",\n", "    allow_flagging=False,\n", "    live=False,\n", "    examples=examples,\n", "    cache_examples=False\n", ")\n", "interface.launch()\n"]}], "metadata": {}, "nbformat": 4, "nbformat_minor": 5}