diff --git a/notebooks/piper_inference_(ONNX).ipynb b/notebooks/piper_inference_(ONNX).ipynb
new file mode 100644
index 0000000..be48807
--- /dev/null
+++ b/notebooks/piper_inference_(ONNX).ipynb
@@ -0,0 +1,475 @@
+{
+ "nbformat": 4,
+ "nbformat_minor": 0,
+ "metadata": {
+ "colab": {
+ "provenance": [],
+ "authorship_tag": "ABX9TyPUlEPUiRIAfFt3VCxArrAD",
+ "include_colab_link": true
+ },
+ "kernelspec": {
+ "name": "python3",
+ "display_name": "Python 3"
+ },
+ "language_info": {
+ "name": "python"
+ }
+ },
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "view-in-github",
+ "colab_type": "text"
+ },
+ "source": [
+ "![\"Open](\"https://colab.research.google.com/assets/colab-badge.svg\")
"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "# **[Piper](https://github.com/rhasspy/piper) inferencing notebook.**\n",
+ "## \n",
+ "\n",
+ "---\n",
+ "\n",
+ "- Notebook made by [rmcpantoja](http://github.com/rmcpantoja)\n",
+ "- Collaborator: [Xx_Nessu_xX](https://fakeyou.com/profile/Xx_Nessu_xX)"
+ ],
+ "metadata": {
+ "id": "eK3nmYDB6C1a"
+ }
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "# First steps"
+ ],
+ "metadata": {
+ "id": "9wIvcSmOby84"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "#@title Install software\n",
+ "#@markdown The speech synthesizer and other important dependencies will be installed in this cell.\n",
+ "\n",
+ "#@markdown #### Enhable Enhanced Accessibility?\n",
+ "#@markdown This Enhanced Accessibility functionality is designed for the visually impaired, in which most of the interface can be used by voice guides.\n",
+ "enhanced_accessibility = True #@param {type:\"boolean\"}\n",
+ "#@markdown ---\n",
+ "\n",
+ "#@markdown #### Please select your language:\n",
+ "lang_select = \"Spanish\" #@param [\"English\", \"Spanish\"]\n",
+ "if lang_select == \"English\":\n",
+ " lang = \"en\"\n",
+ "elif lang_select == \"Spanish\":\n",
+ " lang = \"es\"\n",
+ "else:\n",
+ " raise Exception(\"Language not supported.\")\n",
+ "\n",
+ "if enhanced_accessibility:\n",
+ " from google.colab import output\n",
+ " guideurl = f\"https://github.com/rmcpantoja/piper/blob/master/notebooks/wav/{lang}\"\n",
+ " def playaudio(filename, extension = \"wav\"):\n",
+ " return output.eval_js(f'new Audio(\"{guideurl}/{filename}.{extension}?raw=true\").play()')\n",
+ "\n",
+ "%cd /content\n",
+ "print(\"Installing...\")\n",
+ "if enhanced_accessibility:\n",
+ " playaudio(\"installing\")\n",
+ "!git clone -q https://github.com/rmcpantoja/piper\n",
+ "%cd /content/piper/src/python\n",
+ "!pip install -q -r requirements.txt\n",
+ "!bash build_monotonic_align.sh\n",
+ "!apt-get install -q espeak-ng\n",
+ "import os\n",
+ "if not os.path.exists(\"/content/piper/src/python/lng\"):\n",
+ " !cp -r \"/content/piper/notebooks/lng\" /content/piper/src/python/lng\n",
+ "import sys\n",
+ "sys.path.append('/content/piper/notebooks')\n",
+ "from translator import *\n",
+ "lan = Translator()\n",
+ "\n",
+ "if enhanced_accessibility:\n",
+ " playaudio(\"installed\")\n",
+ "print(\"Success!\")"
+ ],
+ "metadata": {
+ "cellView": "form",
+ "id": "v8b_PEtXb8co"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "#@title Download your exported model\n",
+ "%cd /content/piper/src/python\n",
+ "import os\n",
+ "#@markdown #### ID or link of the voice package (tar.gz format):\n",
+ "package_url_or_id = \"https://github.com/rhasspy/piper/releases/download/v0.0.2/voice-de-thorsten-low.tar.gz\" #@param {type:\"string\"}\n",
+ "#@markdown ---\n",
+ "print(\"Downloading voice package...\")\n",
+ "if enhanced_accessibility:\n",
+ " playaudio(\"downloading\")\n",
+ "if package_url_or_id.startswith(\"1\"):\n",
+ " !gdown -q \"{package_url_or_id}\" -O \"voice_package.tar.gz\"\n",
+ "elif package_url_or_id.startswith(\"https://drive.google.com/file/d/\"):\n",
+ " !gdown -q \"{package_url_or_id}\" -O \"voice_package.tar.gz\" --fuzzy\n",
+ "else:\n",
+ " !wget -q \"{package_url_or_id}\" -O \"voice_package.tar.gz\"\n",
+ "if os.path.exists(\"/content/piper/src/python/voice_package.tar.gz\"):\n",
+ " !tar -xf voice_package.tar.gz\n",
+ " print(\"Voice package downloaded!\")\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"downloaded\")\n",
+ "else:\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"dwnerror\")\n",
+ " raise Exception(lan.translate(lang, \"Model failed to download!\"))"
+ ],
+ "metadata": {
+ "cellView": "form",
+ "id": "ykIYmVXccg6s"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "# inferencing"
+ ],
+ "metadata": {
+ "id": "MRvkYJF6g5FT"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "#@title run inference\n",
+ "#@markdown #### before you enjoy... Some notes!\n",
+ "#@markdown * You can run the cell to download voice packs and download voices you want at any time, even if you run this cell!\n",
+ "#@markdown * When you download a new voice, run this cell again and you will now be able to toggle between all the ones you download. Incredible, right?\n",
+ "\n",
+ "#@markdown Enjoy!!\n",
+ "\n",
+ "%cd /content/piper/src/python\n",
+ "# original: infer_onnx.py\n",
+ "import json\n",
+ "import logging\n",
+ "import math\n",
+ "import sys\n",
+ "from pathlib import Path\n",
+ "\n",
+ "import numpy as np\n",
+ "import onnxruntime\n",
+ "from piper_train.vits.utils import audio_float_to_int16\n",
+ "import glob\n",
+ "import ipywidgets as widgets\n",
+ "from IPython.display import display, Audio, Markdown, clear_output\n",
+ "from espeak_phonemizer import Phonemizer\n",
+ "from piper_train import phonemize\n",
+ "\n",
+ "_LOGGER = logging.getLogger(\"piper_train.infer_onnx\")\n",
+ "\n",
+ "def detect_onnx_models(path):\n",
+ " onnx_models = glob.glob(path + '/*.onnx')\n",
+ " if len(onnx_models) > 1:\n",
+ " return onnx_models\n",
+ " elif len(onnx_models) == 1:\n",
+ " return onnx_models[0]\n",
+ " else:\n",
+ " return None\n",
+ "\n",
+ "\n",
+ "def main():\n",
+ " \"\"\"Main entry point\"\"\"\n",
+ " models_path = \"/content/piper/src/python\"\n",
+ " logging.basicConfig(level=logging.DEBUG)\n",
+ " sess_options = onnxruntime.SessionOptions()\n",
+ " onnx_models = detect_onnx_models(models_path)\n",
+ " speaker_selection = widgets.Dropdown(\n",
+ " options=[],\n",
+ " description=f'{lan.translate(lang, \"Select speaker\")}:',\n",
+ " layout={'visibility': 'hidden'}\n",
+ " )\n",
+ " display(speaker_selection)\n",
+ " if onnx_models is None:\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"novoices\")\n",
+ " raise Exception(lan.translate(lang, \"No downloaded voice packages!\"))\n",
+ " elif isinstance(onnx_models, str):\n",
+ " onnx_model = onnx_models\n",
+ " model, config = load_onnx(onnx_model, sess_options)\n",
+ " if config[\"num_speakers\"] > 1:\n",
+ " speaker_selection.options = config[\"speaker_id_map\"].values()\n",
+ " speaker_selection.layout.visibility = 'visible'\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"multispeaker\")\n",
+ " else:\n",
+ " speaker_selection.layout.visibility = 'hidden'\n",
+ " else:\n",
+ " voice_model_names = []\n",
+ " for current in onnx_models:\n",
+ " voice_struct = current.split(\"/\")[5]\n",
+ " voice_model_names.append(voice_struct)\n",
+ "\n",
+ " selection = widgets.Dropdown(\n",
+ " options=voice_model_names,\n",
+ " description=f'{lan.translate(lang, \"Select voice package\")}:',\n",
+ " )\n",
+ " load_btn = widgets.Button(\n",
+ " description=lan.translate(lang, \"Load it!\")\n",
+ " )\n",
+ " config = None\n",
+ " def load_model(button):\n",
+ " nonlocal config \n",
+ " global onnx_model\n",
+ " nonlocal model\n",
+ " nonlocal models_path\n",
+ " selected_voice = selection.value\n",
+ " onnx_model = f\"{models_path}/{selected_voice}\"\n",
+ " model, config = load_onnx(onnx_model, sess_options)\n",
+ " if config[\"num_speakers\"] > 1:\n",
+ " speaker_selection.options = config[\"speaker_id_map\"].values()\n",
+ " speaker_selection.layout.visibility = 'visible'\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"multispeaker\")\n",
+ " else:\n",
+ " speaker_selection.layout.visibility = 'hidden'\n",
+ "\n",
+ " load_btn.on_click(load_model)\n",
+ " display(selection, load_btn)\n",
+ " speed_slider = widgets.FloatSlider(\n",
+ " value=1,\n",
+ " min=0.25,\n",
+ " max=4,\n",
+ " step=0.1,\n",
+ " description=lan.translate(lang, \"Rate scale\"),\n",
+ " orientation='horizontal',\n",
+ " )\n",
+ " noise_scale_slider = widgets.FloatSlider(\n",
+ " value=0.667,\n",
+ " min=0.25,\n",
+ " max=4,\n",
+ " step=0.1,\n",
+ " description=lan.translate(lang, \"Phoneme noise scale\"),\n",
+ " orientation='horizontal',\n",
+ " )\n",
+ " noise_scale_w_slider = widgets.FloatSlider(\n",
+ " value=1,\n",
+ " min=0.25,\n",
+ " max=4,\n",
+ " step=0.1,\n",
+ " description=lan.translate(lang, \"Phoneme stressing scale\"),\n",
+ " orientation='horizontal',\n",
+ " )\n",
+ " play = widgets.Checkbox(\n",
+ " value=True,\n",
+ " description=lan.translate(lang, \"Auto-play\"),\n",
+ " disabled=False\n",
+ " )\n",
+ " text_input = widgets.Text(\n",
+ " value='',\n",
+ " placeholder=f'{lan.translate(lang, \"Enter your text here\")}:',\n",
+ " description=lan.translate(lang, \"Text to synthesize\"),\n",
+ " layout=widgets.Layout(width='80%')\n",
+ " )\n",
+ " synthesize_button = widgets.Button(\n",
+ " description=lan.translate(lang, \"Synthesize\"),\n",
+ " button_style='success', # 'success', 'info', 'warning', 'danger' or ''\n",
+ " tooltip=lan.translate(lang, \"Click here to synthesize the text.\"),\n",
+ " icon='check'\n",
+ " )\n",
+ " close_button = widgets.Button(\n",
+ " description=lan.translate(lang, \"Exit\"),\n",
+ " tooltip=lan.translate(lang, \"Closes this GUI.\"),\n",
+ " icon='check'\n",
+ " )\n",
+ "\n",
+ " def on_synthesize_button_clicked(b):\n",
+ " if model is None:\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"nomodel\")\n",
+ " raise Exception(lan.translate(lang, \"You have not loaded any model from the list!\"))\n",
+ " text = text_input.value\n",
+ " if config[\"num_speakers\"] > 1:\n",
+ " sid = speaker_selection.value\n",
+ " else:\n",
+ " sid = None\n",
+ " rate = speed_slider.value\n",
+ " noise_scale = noise_scale_slider.value\n",
+ " noise_scale_w = noise_scale_w_slider.value\n",
+ " auto_play = play.value\n",
+ " inferencing(model, config, sid, text, rate, noise_scale, noise_scale_w, auto_play)\n",
+ " \n",
+ " def on_close_button_clicked(b):\n",
+ " clear_output()\n",
+ " if enhanced_accessibility:\n",
+ " playaudio(\"exit\")\n",
+ " \n",
+ " synthesize_button.on_click(on_synthesize_button_clicked)\n",
+ " close_button.on_click(on_close_button_clicked)\n",
+ " display(text_input)\n",
+ " display(speed_slider)\n",
+ " display(noise_scale_slider)\n",
+ " display(noise_scale_w_slider)\n",
+ " display(play)\n",
+ " display(synthesize_button)\n",
+ " display(close_button)\n",
+ "\n",
+ "def load_onnx(model, sess_options):\n",
+ " _LOGGER.debug(\"Loading model from %s\", model)\n",
+ " config = load_config(model)\n",
+ " model = onnxruntime.InferenceSession(str(model), sess_options=sess_options)\n",
+ " _LOGGER.info(\"Loaded model from %s\", model)\n",
+ " return model, config\n",
+ "\n",
+ "def load_config(model):\n",
+ " with open(f\"{model}.json\", \"r\") as file:\n",
+ " config = json.load(file)\n",
+ " return config\n",
+ "\n",
+ "def inferencing(model, config, sid, line, length_scale = 1, noise_scale = 0.667, noise_scale_w = 0.8, auto_play=True):\n",
+ " espeak_voice = config[\"espeak\"][\"voice\"]\n",
+ " phonemizer = Phonemizer(default_voice=espeak_voice)\n",
+ " phonemes = phonemize.phonemize(line, phonemizer)\n",
+ " ids = phonemize.phonemes_to_ids(phonemes)\n",
+ " phoneme_ids = ids\n",
+ " num_speakers = config[\"num_speakers\"]\n",
+ " if num_speakers == 1:\n",
+ " speaker_id = None # for now\n",
+ " else:\n",
+ " speaker_id = sid\n",
+ " text = np.expand_dims(np.array(phoneme_ids, dtype=np.int64), 0)\n",
+ " text_lengths = np.array([text.shape[1]], dtype=np.int64)\n",
+ " scales = np.array(\n",
+ " [noise_scale, length_scale, noise_scale_w],\n",
+ " dtype=np.float32,\n",
+ " )\n",
+ " sid = None\n",
+ " if speaker_id is not None:\n",
+ " sid = np.array([speaker_id], dtype=np.int64)\n",
+ " audio = model.run(\n",
+ " None,\n",
+ " {\n",
+ " \"input\": text,\n",
+ " \"input_lengths\": text_lengths,\n",
+ " \"scales\": scales,\n",
+ " \"sid\": sid,\n",
+ " },\n",
+ " )[0].squeeze((0, 1))\n",
+ " audio = audio_float_to_int16(audio.squeeze())\n",
+ " sample_rate = config[\"audio\"][\"sample_rate\"]\n",
+ " display(Markdown(f\"{line}\"))\n",
+ " display(Audio(audio, rate=sample_rate, autoplay=auto_play))\n",
+ "\n",
+ "def denoise(\n",
+ " audio: np.ndarray, bias_spec: np.ndarray, denoiser_strength: float\n",
+ ") -> np.ndarray:\n",
+ " audio_spec, audio_angles = transform(audio)\n",
+ "\n",
+ " a = bias_spec.shape[-1]\n",
+ " b = audio_spec.shape[-1]\n",
+ " repeats = max(1, math.ceil(b / a))\n",
+ " bias_spec_repeat = np.repeat(bias_spec, repeats, axis=-1)[..., :b]\n",
+ "\n",
+ " audio_spec_denoised = audio_spec - (bias_spec_repeat * denoiser_strength)\n",
+ " audio_spec_denoised = np.clip(audio_spec_denoised, a_min=0.0, a_max=None)\n",
+ " audio_denoised = inverse(audio_spec_denoised, audio_angles)\n",
+ "\n",
+ " return audio_denoised\n",
+ "\n",
+ "\n",
+ "def stft(x, fft_size, hopsamp):\n",
+ " \"\"\"Compute and return the STFT of the supplied time domain signal x.\n",
+ " Args:\n",
+ " x (1-dim Numpy array): A time domain signal.\n",
+ " fft_size (int): FFT size. Should be a power of 2, otherwise DFT will be used.\n",
+ " hopsamp (int):\n",
+ " Returns:\n",
+ " The STFT. The rows are the time slices and columns are the frequency bins.\n",
+ " \"\"\"\n",
+ " window = np.hanning(fft_size)\n",
+ " fft_size = int(fft_size)\n",
+ " hopsamp = int(hopsamp)\n",
+ " return np.array(\n",
+ " [\n",
+ " np.fft.rfft(window * x[i : i + fft_size])\n",
+ " for i in range(0, len(x) - fft_size, hopsamp)\n",
+ " ]\n",
+ " )\n",
+ "\n",
+ "\n",
+ "def istft(X, fft_size, hopsamp):\n",
+ " \"\"\"Invert a STFT into a time domain signal.\n",
+ " Args:\n",
+ " X (2-dim Numpy array): Input spectrogram. The rows are the time slices and columns are the frequency bins.\n",
+ " fft_size (int):\n",
+ " hopsamp (int): The hop size, in samples.\n",
+ " Returns:\n",
+ " The inverse STFT.\n",
+ " \"\"\"\n",
+ " fft_size = int(fft_size)\n",
+ " hopsamp = int(hopsamp)\n",
+ " window = np.hanning(fft_size)\n",
+ " time_slices = X.shape[0]\n",
+ " len_samples = int(time_slices * hopsamp + fft_size)\n",
+ " x = np.zeros(len_samples)\n",
+ " for n, i in enumerate(range(0, len(x) - fft_size, hopsamp)):\n",
+ " x[i : i + fft_size] += window * np.real(np.fft.irfft(X[n]))\n",
+ " return x\n",
+ "\n",
+ "\n",
+ "def inverse(magnitude, phase):\n",
+ " recombine_magnitude_phase = np.concatenate(\n",
+ " [magnitude * np.cos(phase), magnitude * np.sin(phase)], axis=1\n",
+ " )\n",
+ "\n",
+ " x_org = recombine_magnitude_phase\n",
+ " n_b, n_f, n_t = x_org.shape # pylint: disable=unpacking-non-sequence\n",
+ " x = np.empty([n_b, n_f // 2, n_t], dtype=np.complex64)\n",
+ " x.real = x_org[:, : n_f // 2]\n",
+ " x.imag = x_org[:, n_f // 2 :]\n",
+ " inverse_transform = []\n",
+ " for y in x:\n",
+ " y_ = istft(y.T, fft_size=1024, hopsamp=256)\n",
+ " inverse_transform.append(y_[None, :])\n",
+ "\n",
+ " inverse_transform = np.concatenate(inverse_transform, 0)\n",
+ "\n",
+ " return inverse_transform\n",
+ "\n",
+ "\n",
+ "def transform(input_data):\n",
+ " x = input_data\n",
+ " real_part = []\n",
+ " imag_part = []\n",
+ " for y in x:\n",
+ " y_ = stft(y, fft_size=1024, hopsamp=256).T\n",
+ " real_part.append(y_.real[None, :, :]) # pylint: disable=unsubscriptable-object\n",
+ " imag_part.append(y_.imag[None, :, :]) # pylint: disable=unsubscriptable-object\n",
+ " real_part = np.concatenate(real_part, 0)\n",
+ " imag_part = np.concatenate(imag_part, 0)\n",
+ "\n",
+ " magnitude = np.sqrt(real_part**2 + imag_part**2)\n",
+ " phase = np.arctan2(imag_part.data, real_part.data)\n",
+ "\n",
+ " return magnitude, phase\n",
+ "\n",
+ "main()"
+ ],
+ "metadata": {
+ "id": "hcKk8M2ug8kM",
+ "cellView": "form"
+ },
+ "execution_count": null,
+ "outputs": []
+ }
+ ]
+}
\ No newline at end of file