docs/tutorial/tts/tts_tutorial.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a href=\"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/PaddlePaddle/PaddleSpeech\"><img style=\"position: absolute; z-index: 999; top: 0; right: 0; border: 0; width: 128px; height: 128px;\" src=\"https://nosir.github.io/cleave.js/images/right-graphite@2x.png\" alt=\"Fork me on GitHub\"></a>\n",
    "\n",
    "# 『听』和『说』\n",
    "人类通过听觉获取的信息大约占所有感知信息的 20% ~ 30%。声音存储了丰富的语义以及时序信息，由专门负责听觉的器官接收信号，产生一系列连锁刺激后，在人类大脑的皮层听区进行处理分析，获取语义和知识。近年来，随着深度学习算法上的进步以及不断丰厚的硬件资源条件，**文本转语音（Text-to-Speech, TTS）** 技术在移动、虚拟娱乐等领域得到了广泛的应用。</font>\n",
    "## \"听\"书\n",
    "使用 [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR) 直接获取书籍上的文字。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# download demo sources\n",
    "!mkdir download\n",
    "!wget -P download https://paddlespeech.bj.bcebos.com/tutorial/tts/ocr_result.jpg\n",
    "!wget -P download https://paddlespeech.bj.bcebos.com/tutorial/tts/ocr.wav\n",
    "!wget -P download https://paddlespeech.bj.bcebos.com/tutorial/tts/tts_lips.mp4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import IPython.display as dp\n",
    "from PIL import Image\n",
    "img_path = 'download/ocr_result.jpg'\n",
    "im = Image.open(img_path)\n",
    "dp.display(im)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "使用 [PaddleSpeech](https://github.com/PaddlePaddle/PaddleSpeech)，阅读上一步识别出来的文字。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dp.Audio(\"download/ocr.wav\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "具体实现代码详见 [Story Talker](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/demos/story_talker)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 偶像开口说话\n",
    "*元宇宙来袭，构造你的虚拟人！* 看看 [PaddleGAN](https://github.com/PaddlePaddle/PaddleGAN) 怎样合成唇形，让WiFi之母——海蒂·拉玛说话。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from IPython.display import HTML\n",
    "html_str = '''\n",
    "<video controls width=\"600\" height=\"360\" src=\"{}\">animation</video>\n",
    "'''.format(\"download/tts_lips.mp4\")\n",
    "dp.display(HTML(html_str))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "具体实现代码请参考 [Metaverse](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/demos/metaverse)。\n",
    "\n",
    "下面让我们来系统地学习语音方面的知识，看看怎样使用 **PaddleSpeech** 实现基本的语音功能，以及怎样结合光学字符识别（Optical Character Recognition，OCR）、自然语言处理（Natural Language Processing，NLP）等技术“听”书、让名人开口说话。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 前言\n",
    "## 背景知识\n",
    "为了更好地了解文本转语音任务的要素，我们先简要地回顾一下文本转语音的发展历史。如果你对此已经有所了解，或希望能尽快使用代码实现，请直接跳至[实践](#实践)。\n",
    "### 定义\n",
    "<!----\n",
    "Note: \n",
    "1.此句抄自 [李沐Dive into Dive Learning](https://zh-v2.d2l.ai/chapter_introduction/index.html)\n",
    "2.修改参考A survey on Neural Speech Sysnthesis.\n",
    "---> \n",
    "文本转语音，又称语音合成（Speech Sysnthesis），指的是将一段文本按照一定需求转化成对应的音频，这种特性决定了的输出数据比输入输入长得多。文本转语音是一项包含了语义学、声学、数字信号处理以及机器学习的等多项学科的交叉任务。虽然辨识低质量音频文件的内容对人类来说很容易，但这对计算机来说并非易事。\n",
    "\n",
    "按照不同的应用需求，更广义的语音合成研究包括：*语音转换*，例如说话人转换、语音到歌唱转换、语音情感转换、口音转换等；*歌唱合成*，例如歌词到歌唱转换、可视语音合成等。\n",
    "\n",
    "### 发展历史\n",
    "\n",
    "<!--\n",
    "以下摘自维基百科 https://en.wikipedia.org/wiki/Speech_synthesis\n",
    "--->\n",
    "\n",
    "在第二次工业革命之前，语音的合成主要以机械式的音素合成为主。1779年，德裔丹麦科学家 Christian Gottlieb Kratzenstein 建造了人类的声道模型，使其可以产生五个长元音。1791年， Wolfgang von Kempelen 添加了唇和舌的模型，使其能够发出辅音和元音。贝尔实验室于20世纪30年代发明了声码器（Vocoder），将语音自动分解为音调和共振，此项技术由 Homer Dudley 改进为键盘式合成器并于 1939年纽约世界博览会展出。\n",
    "\n",
    "第一台基于计算机的语音合成系统起源于20世纪50年代。1961年，IBM 的 John Larry Kelly，以及 Louis Gerstman 使用 IBM 704 计算机合成语音，成为贝尔实验室最著名的成就之一。1975年，第一代语音合成系统之一 —— MUSA（MUltichannel Speaking Automation）问世，其由一个独立的硬件和配套的软件组成。1978年发行的第二个版本也可以进行无伴奏演唱。90 年代的主流是采用 MIT 和贝尔实验室的系统，并结合自然语言处理模型。\n",
    "\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/55035de353b042cd8c4468819b2d36e2fcc89bffdf2b442fa4c7b0b5499e1592\"></center>\n",
    "\n",
    "### 主流方法\n",
    "\n",
    "当前的主流方法分为**基于统计参数的语音合成**、**波形拼接语音合成**、**混合方法**以及**端到端神经网络语音合成**。基于参数的语音合成包含隐马尔可夫模型（Hidden Markov Model,HMM）以及深度学习网络（Deep Neural Network，DNN）。端到端的方法保函声学模型+声码器以及“完全”端到端方法。\n",
    "\n",
    "\n",
    "## 基于深度学习的语音合成技术\n",
    "\n",
    "### 语音合成基本知识\n",
    "\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/10859679d74745ab82fb6f5c9984a95152c25b0e3dce4515b120c8997a6752d8\"></center>\n",
    "<br></br>\n",
    "\n",
    "语音合成流水线包含 <font color=\"#ff0000\">**文本前端（Text Frontend）**</font> 、<font color=\"#ff0000\">**声学模型（Acoustic Model）**</font> 和 <font color=\"#ff0000\">**声码器（Vocoder）**</font> 三个主要模块:\n",
    "- 通过文本前端模块将原始文本转换为字符/音素。\n",
    "- 通过声学模型将字符/音素转换为声学特征，如线性频谱图、mel 频谱图、LPC 特征等。\n",
    "- 通过声码器将声学特征转换为波形。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 实践"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 安装 paddlespeech"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install --upgrade pip && pip install paddlespeech -U"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "环境安装请参考 [Installation](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/install.md) 教程。 \n",
    "下面使用 **PaddleSpeech** 提供的预训练模型合成中文语音。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 数据及模型准备"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 获取PaddlePaddle预训练模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!wget -P download https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwg_baker_ckpt_0.4.zip\n",
    "!unzip -d download download/pwg_baker_ckpt_0.4.zip\n",
    "!wget -P download https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_nosil_baker_ckpt_0.4.zip\n",
    "!unzip -d download download/fastspeech2_nosil_baker_ckpt_0.4.zip"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!tree download/pwg_baker_ckpt_0.4\n",
    "!tree download/fastspeech2_nosil_baker_ckpt_0.4"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 导入 Python 包"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 本项目的依赖需要用到 nltk 包，但是有时会因为网络原因导致不好下载，此处手动下载一下放到百度服务器的包\n",
    "!wget https://paddlespeech.bj.bcebos.com/Parakeet/tools/nltk_data.tar.gz\n",
    "!tar zxvf nltk_data.tar.gz"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 设置 gpu 环境\n",
    "%env CUDA_VISIBLE_DEVICES=0\n",
    "\n",
    "import logging\n",
    "import sys\n",
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\", category=DeprecationWarning)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import argparse\n",
    "import os\n",
    "from pathlib import Path\n",
    "import IPython.display as dp\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import paddle\n",
    "import soundfile as sf\n",
    "import yaml\n",
    "from paddlespeech.t2s.frontend.zh_frontend import Frontend\n",
    "from paddlespeech.t2s.models.fastspeech2 import FastSpeech2\n",
    "from paddlespeech.t2s.models.fastspeech2 import FastSpeech2Inference\n",
    "from paddlespeech.t2s.models.parallel_wavegan import PWGGenerator\n",
    "from paddlespeech.t2s.models.parallel_wavegan import PWGInference\n",
    "from paddlespeech.t2s.modules.normalizer import ZScore\n",
    "from yacs.config import CfgNode"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 设置预训练模型的路径"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "fastspeech2_config = \"download/fastspeech2_nosil_baker_ckpt_0.4/default.yaml\"\n",
    "fastspeech2_checkpoint = \"download/fastspeech2_nosil_baker_ckpt_0.4/snapshot_iter_76000.pdz\"\n",
    "fastspeech2_stat = \"download/fastspeech2_nosil_baker_ckpt_0.4/speech_stats.npy\"\n",
    "pwg_config = \"download/pwg_baker_ckpt_0.4/pwg_default.yaml\"\n",
    "pwg_checkpoint = \"download/pwg_baker_ckpt_0.4/pwg_snapshot_iter_400000.pdz\"\n",
    "pwg_stat = \"download/pwg_baker_ckpt_0.4/pwg_stats.npy\"\n",
    "phones_dict = \"download/fastspeech2_nosil_baker_ckpt_0.4/phone_id_map.txt\"\n",
    "# 读取 conf 配置文件并结构化\n",
    "with open(fastspeech2_config) as f:\n",
    "    fastspeech2_config = CfgNode(yaml.safe_load(f))\n",
    "with open(pwg_config) as f:\n",
    "    pwg_config = CfgNode(yaml.safe_load(f))\n",
    "print(\"========Config========\")\n",
    "print(fastspeech2_config)\n",
    "print(\"---------------------\")\n",
    "print(pwg_config)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 文本前端（Text Frontend）\n",
    "\n",
    "一个文本前端模块主要包含:\n",
    "- 分段（Text Segmentation）\n",
    "- 文本正则化（Text Normalization, TN）\n",
    "- 分词（Word Segmentation, 主要是在中文中）\n",
    "- 词性标注（Part-of-Speech, PoS）\n",
    "- 韵律预测（Prosody）\n",
    "- 字音转换（Grapheme-to-Phoneme，G2P）\n",
    "<font size=2>（Grapheme: **语言**书写系统的最小有意义单位; Phoneme: 区分单词的最小**语音**单位）</font>\n",
    "    - 多音字（Polyphone）\n",
    "    - 变调（Tone Sandhi）\n",
    "        - “一”、“不”变\n",
    "        - 三声变调\n",
    "        - 轻声变调\n",
    "        - 儿化音\n",
    "        - 方言\n",
    "- ...\n",
    "\n",
    "（输入给声学模型之前，还需要把音素序列转换为 id）\n",
    "\n",
    "\n",
    "其中最重要的模块是<font color=\"#ff0000\"> 文本正则化 </font>模块和<font color=\"#ff0000\"> 字音转换（TTS 中更常用 G2P 代指） </font>模块。\n",
    "\n",
    "\n",
    "各模块输出示例:\n",
    "```text\n",
    "• Text: 全国一共有112所211高校\n",
    "• Text Normalization: 全国一共有一百一十二所二一一高校\n",
    "• Word Segmentation: 全国/一共/有/一百一十二/所/二一一/高校/\n",
    "• G2P（注意此句中“一”的读音）:\n",
    "    quan2 guo2 yi2 gong4 you3 yi4 bai3 yi1 shi2 er4 suo3 er4 yao1 yao1 gao1 xiao4\n",
    "    （可以进一步把声母和韵母分开）\n",
    "    q uan2 g uo2 y i2 g ong4 y ou3 y i4 b ai3 y i1 sh i2 er4 s uo3 er4 y ao1 y ao1 g ao1 x iao4\n",
    "    （把音调和声韵母分开）\n",
    "    q uan g uo y i g ong y ou y i b ai y i sh i er s uo er y ao y ao g ao x iao\n",
    "    0 2 0 2 0 2 0 4 0 3 ...\n",
    "• Prosody (prosodic words #1, prosodic phrases #2, intonation phrases #3, sentence #4):\n",
    "    全国#2一共有#2一百#1一十二所#2二一一#1高校#4\n",
    "    （分词的结果一般是固定的，但是不同人习惯不同，可能有不同的韵律）\n",
    "```\n",
    "\n",
    "文本前端模块的设计需要结合很多专业的语义学知识和经验。人类在读文本的时候可以自然而然地读出正确的发音，但是这些先验知识计算机并不知晓。\n",
    "例如，对于一个句子的分词：\n",
    "\n",
    "```text\n",
    "我也想过过过儿过过的生活\n",
    "我也想/过过/过儿/过过的/生活\n",
    "\n",
    "货拉拉拉不拉拉布拉多\n",
    "货拉拉/拉不拉/拉布拉多\n",
    "\n",
    "南京市长江大桥\n",
    "南京市长/江大桥\n",
    "南京市/长江大桥\n",
    "```\n",
    "或者是词的变调和儿化音：\n",
    "```\n",
    "你要不要和我们一起出去玩？\n",
    "你要不（2声）要和我们一（4声）起出去玩（儿）？\n",
    "\n",
    "不好，我要一个人出去。\n",
    "不（4声）好，我要一（2声）个人出去。\n",
    "\n",
    "（以下每个词的所有字都是三声的，请你读一读，体会一下在读的时候，是否每个字都被读成了三声？）\n",
    "纸老虎、虎骨酒、展览馆、岂有此理、手表厂有五种好产品\n",
    "```\n",
    "又或是多音字，这类情况通常需要先正确分词：\n",
    "```text\n",
    "人要行，干一行行一行，一行行行行行;\n",
    "人要是不行，干一行不行一行，一行不行行行不行。\n",
    "\n",
    "佟大为妻子产下一女\n",
    "\n",
    "海水朝朝朝朝朝朝朝落\n",
    "浮云长长长长长长长消\n",
    "```\n",
    "\n",
    "PaddleSpeech Text-to-Speech的文本前端解决方案:\n",
    "- [文本正则](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/tn)\n",
    "- [G2P](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/g2p)\n",
    "    - 多音字模块: pypinyin/g2pM\n",
    "    - 变调模块: 用分词 + 规则"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 构造文本前端对象\n",
    "传入`phones_dict`，把相应的`phones`转换成`phone_ids`。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 传入 phones_dict 会把相应的 phones 转换成 phone_ids\n",
    "frontend = Frontend(phone_vocab_path=phones_dict)\n",
    "print(\"Frontend done!\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 调用文本前端\n",
    "\n",
    "文本前端对输入数据进行正则化时会进行分句，若`merge_sentences`设置为`False`，则所有分句的 `phone_ids` 构成一个 `List`；若设置为`True`，`input_ids[\"phone_ids\"][0]`则表示整句的`phone_ids`。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# input = \"我每天中午12:00起床\"\n",
    "# input = \"我出生于2005/11/08，那天的最低气温达到-10°C\"\n",
    "input = \"你好，欢迎使用百度飞桨框架进行深度学习研究！\"\n",
    "input_ids = frontend.get_input_ids(input, merge_sentences=True, print_info=True)\n",
    "phone_ids = input_ids[\"phone_ids\"][0]\n",
    "print(\"phone_ids:%s\"%phone_ids)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 用深度学习实现文本前端\n",
    "\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/85a5cd8aef1e444cbb980a2f1f184316247bbb7870a34925a77b799802df8ef0\"></center>\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 声学模型（Acoustic Model）\n",
    "\n",
    "声学模型将字符/音素转换为声学特征，如线性频谱图、mel 频谱图、LPC 特征等，声学特征以 “帧” 为单位，一般一帧是 10ms 左右，一个音素一般对应 5~20 帧左右, 声学模型需要解决的是 <font color=\"#ff0000\">“不等长序列间的映射问题”</font>，“不等长”是指，同一个人发不同音素的持续时间不同，同一个人在不同时刻说同一句话的语速可能不同，对应各个音素的持续时间不同，不同人说话的特色不同，对应各个音素的持续时间不同。这是一个困难的“一对多”问题。\n",
    "```\n",
    "# 卡尔普陪外孙玩滑梯\n",
    "000001|baker_corpus|sil 20 k 12 a2 4 er2 10 p 12 u3 12 p 9 ei2 9 uai4 15 s 11 uen1 12 uan2 14 h 10 ua2 11 t 15 i1 16 sil 20\n",
    "```\n",
    "\n",
    "声学模型主要分为自回归模型和非自回归模型，其中自回归模型在 `t` 时刻的预测需要依赖 `t-1` 时刻的输出作为输入，预测时间长，但是音质相对较好，非自回归模型不存在预测上的依赖关系，预测时间快，音质相对较差。\n",
    "\n",
    "主流声学模型发展的脉络:\n",
    "- 自回归模型:\n",
    "    - Tacotron\n",
    "    - Tacotron2\n",
    "    - Transformer TTS\n",
    "- 非自回归模型:\n",
    "    - FastSpeech\n",
    "    - SpeedySpeech\n",
    "    - FastPitch\n",
    "    - FastSpeech2\n",
    "    - ...\n",
    " \n",
    "在本教程中，我们使用 `FastSpeech2` 作为声学模型。\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/6b6d671713ec4d20a0e60653c7a5d4ae3c35b1d1e58b4cc39e0bc82ad4a341d9\"></center>\n",
    "<br><center> FastSpeech2 网络结构图</center></br>\n",
    "\n",
    "\n",
    "PaddleSpeech TTS 实现的 FastSpeech2 与论文不同的地方在于，我们使用的是 phone 级别的 `pitch` 和 `energy`(与 FastPitch 类似)，这样的合成结果可以更加**稳定**。\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/862c21456c784c41a83a308b7d9707f0810cc3b3c6f94ed48c60f5d32d0072f0\"></center>\n",
    "<br><center> FastPitch 网络结构图</center></br>\n",
    "\n",
    "更多关于[语音合成模型的发展及改进](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/tts/models_introduction.md)。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 初始化声学模型 FastSpeech2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(phones_dict, \"r\") as f:\n",
    "    phn_id = [line.strip().split() for line in f.readlines()]\n",
    "vocab_size = len(phn_id)\n",
    "print(\"vocab_size:\", vocab_size)\n",
    "odim = fastspeech2_config.n_mels\n",
    "model = FastSpeech2(\n",
    "    idim=vocab_size, odim=odim, **fastspeech2_config[\"model\"])\n",
    "# 加载预训练模型参数\n",
    "model.set_state_dict(paddle.load(fastspeech2_checkpoint)[\"main_params\"])\n",
    "# 推理阶段不启用 batch norm 和 dropout\n",
    "model.eval()\n",
    "stat = np.load(fastspeech2_stat)\n",
    "# 读取数据预处理阶段数据集的均值和标准差\n",
    "mu, std = stat\n",
    "mu, std = paddle.to_tensor(mu), paddle.to_tensor(std)\n",
    "# 构造归一化的新模型\n",
    "fastspeech2_normalizer = ZScore(mu, std)\n",
    "fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model)\n",
    "fastspeech2_inference.eval()\n",
    "print(\"FastSpeech2 done!\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 调用声学模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with paddle.no_grad():\n",
    "    mel = fastspeech2_inference(phone_ids)\n",
    "print(\"shepe of mel (n_frames x n_mels):\")\n",
    "print(mel.shape)\n",
    "# 绘制声学模型输出的 mel 频谱\n",
    "fig, ax = plt.subplots(figsize=(16, 6))\n",
    "im = ax.imshow(mel.T, aspect='auto',origin='lower')\n",
    "plt.title('Mel Spectrogram')\n",
    "plt.xlabel('Time')\n",
    "plt.ylabel('Frequency')\n",
    "plt.tight_layout()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 声码器（Vocoder）\n",
    "声码器将声学特征转换为波形。声码器需要解决的是 <font color=\"#ff0000\">“信息缺失的补全问题”</font>。信息缺失是指，在音频波形转换为频谱图的时候，存在**相位信息**的缺失，在频谱图转换为 mel 频谱图的时候，存在**频域压缩**导致的信息缺失；假设音频的采样率是16kHZ, 一帧的音频有 10ms，也就是说，1s 的音频有 16000 个采样点，而 1s 中包含 100 帧，每一帧有 160 个采样点，声码器的作用就是将一个频谱帧变成音频波形的 160 个采样点，所以声码器中一般会包含**上采样**模块。\n",
    "\n",
    "与声学模型类似，声码器也分为自回归模型和非自回归模型, 更细致的分类如下:\n",
    "\n",
    "- Autoregression\n",
    "    - WaveNet\n",
    "    - WaveRNN\n",
    "    - LPCNet\n",
    "- Flow\n",
    "    - <font color=\"#ff0000\">WaveFlow</font>\n",
    "    - WaveGlow\n",
    "    - FloWaveNet\n",
    "    - Parallel WaveNet\n",
    "- GAN\n",
    "\t- WaveGAN\n",
    "    - <font color=\"#ff0000\">Parallel WaveGAN</font>\n",
    "    - <font color=\"#ff0000\">MelGAN</font>\n",
    "    - <font color=\"#ff0000\">Style MelGAN</font>\n",
    "    - <font color=\"#ff0000\">Multi Band MelGAN</font>\n",
    "    - <font color=\"#ff0000\">HiFi GAN</font>\n",
    "- VAE\n",
    "    - Wave-VAE\n",
    "- Diffusion\n",
    "    - WaveGrad\n",
    "    - DiffWave\n",
    "\n",
    "PaddleSpeech TTS 主要实现了百度的 `WaveFlow` 和一些主流的 GAN Vocoder, 在本教程中，我们使用 `Parallel WaveGAN` 作为声码器。\n",
    "\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/9eafa4e5642d45309e6e8883bff46380407b3858d0934bf5896868281316ce94\" width=\"700\"></center>\n",
    "<br><center>图1：Parallel WaveGAN 网络结构图</center></br>\n",
    "\n",
    "各 GAN Vocoder 的生成器和判别器的 Loss 的区别如下表格所示:\n",
    "\n",
    "Model  | Generator Loss |Discriminator Loss\n",
    ":-------------:| :------------:| :-----\n",
    "Mel GAN| adversial loss <br> Feature Matching  | Multi-Scale Discriminator |\n",
    "Parallel Wave GAN|adversial loss <br> Multi-resolution STFT loss  | adversial loss|\n",
    "Multi-Band Mel GAN | adversial loss <br> full band Multi-resolution STFT loss <br> sub band Multi-resolution STFT loss |Multi-Scale Discriminator|\n",
    "HiFi GAN |adversial loss <br> Feature Matching <br>  Mel-Spectrogram Loss | Multi-Scale Discriminator <br> Multi-Period Discriminator| \n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 初始化声码器 Parallel WaveGAN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "vocoder = PWGGenerator(**pwg_config[\"generator_params\"])\n",
    "# 模型加载预训练参数\n",
    "vocoder.set_state_dict(paddle.load(pwg_checkpoint)[\"generator_params\"])\n",
    "vocoder.remove_weight_norm()\n",
    "# 推理阶段不启用 batch norm 和 dropout\n",
    "vocoder.eval()\n",
    "# 读取数据预处理阶段数据集的均值和标准差\n",
    "stat = np.load(pwg_stat)\n",
    "mu, std = stat\n",
    "mu, std = paddle.to_tensor(mu), paddle.to_tensor(std)\n",
    "pwg_normalizer = ZScore(mu, std)\n",
    "# 构建归一化的模型\n",
    "pwg_inference = PWGInference(pwg_normalizer, vocoder)\n",
    "pwg_inference.eval()\n",
    "print(\"Parallel WaveGAN done!\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 调用声码器"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with paddle.no_grad():\n",
    "    wav = pwg_inference(mel)\n",
    "print(\"shepe of wav (time x n_channels):%s\"%wav.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 绘制声码器输出的波形图\n",
    "wave_data = wav.numpy().T\n",
    "time = np.arange(0, wave_data.shape[1]) * (1.0 / fastspeech2_config.fs)\n",
    "fig, ax = plt.subplots(figsize=(16, 6))\n",
    "plt.plot(time, wave_data[0])\n",
    "plt.title('Waveform')\n",
    "plt.xlabel('Time (seconds)')\n",
    "plt.ylabel('Amplitude (normed)')\n",
    "plt.tight_layout()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 播放音频"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dp.Audio(wav.numpy().T, rate=fastspeech2_config.fs)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 保存音频"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!mkdir output\n",
    "sf.write(\n",
    "    \"output/output.wav\",\n",
    "    wav.numpy(),\n",
    "    samplerate=fastspeech2_config.fs)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 进阶 —— 个性化调节\n",
    "FastSpeech2 模型可以个性化地调节音素时长、音调和能量，通过一些简单的调节就可以获得一些有意思的效果。\n",
    "\n",
    "例如对于以下的原始音频`\"凯莫瑞安联合体的经济崩溃，迫在眉睫\"`。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 原始音频\n",
    "dp.display(dp.Audio(url=\"https://paddlespeech.bj.bcebos.com/Parakeet/docs/demos/speed/x1_001.wav\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# speed x 1.2\n",
    "dp.display(dp.Audio(url=\"https://paddlespeech.bj.bcebos.com/Parakeet/docs/demos/speed/x1.2_001.wav\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# speed x 0.8\n",
    "dp.display(dp.Audio(url=\"https://paddlespeech.bj.bcebos.com/Parakeet/docs/demos/speed/x0.8_001.wav\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# pitch x 1.3(童声)\n",
    "dp.display(dp.Audio(url=\"https://paddlespeech.bj.bcebos.com/Parakeet/docs/demos/child_voice/001.wav\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# robot\n",
    "dp.display(dp.Audio(url=\"https://paddlespeech.bj.bcebos.com/Parakeet/docs/demos/robot/001.wav\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "具体实现代码请参考 [Style FastSpeech2](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/demos/style_fs2)。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 用 PaddleSpeech 训练 TTS 模型\n",
    "PaddleSpeech 的 examples 是按照 数据集/模型 的结构安排的:\n",
    "```text\n",
    "examples   \n",
    "├── aishell3\n",
    "│   ├── README.md\n",
    "│   ├── tts3\n",
    "│   └── vc0\n",
    "├── csmsc\n",
    "│   ├── README.md\n",
    "│   ├── tts2\n",
    "│   ├── tts3\n",
    "│   ├── voc1\n",
    "│   └── voc3\n",
    "├── ...\n",
    "└── ...\n",
    "```\n",
    "我们在每个数据集的 README.md 介绍了子目录和模型的对应关系, 在 TTS 中有如下对应关系:\n",
    "```text\n",
    "tts0 - Tacotron2\n",
    "tts1 - TransformerTTS\n",
    "tts2 - SpeedySpeech\n",
    "tts3 - FastSpeech2\n",
    "voc0 - WaveFlow\n",
    "voc1 - Parallel WaveGAN\n",
    "voc2 - MelGAN\n",
    "voc3 - MultiBand MelGAN\n",
    "```\n",
    "### 基于 CSMCS 数据集训练 FastSpeech2 模型\n",
    "```bash\n",
    "git clone https://github.com/PaddlePaddle/PaddleSpeech.git\n",
    "cd examples/csmsc/tts3\n",
    "```\n",
    "根据 README.md, 下载 CSMCS 数据集和其对应的强制对齐文件, 并放置在对应的位置\n",
    "```bash\n",
    "./run.sh\n",
    "```\n",
    "`run.sh` 中包含预处理、训练、合成、静态图推理等步骤:\n",
    "\n",
    "```bash\n",
    "#!/bin/bash\n",
    "set -e\n",
    "source path.sh\n",
    "gpus=0,1\n",
    "stage=0\n",
    "stop_stage=100\n",
    "conf_path=conf/default.yaml\n",
    "train_output_path=exp/default\n",
    "ckpt_name=snapshot_iter_153.pdz\n",
    "\n",
    "# with the following command, you can choice the stage range you want to run\n",
    "# such as `./run.sh --stage 0 --stop-stage 0`\n",
    "# this can not be mixed use with `$1`, `$2` ...\n",
    "source ${MAIN_ROOT}/utils/parse_options.sh || exit 1\n",
    "\n",
    "if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then\n",
    "    # prepare data\n",
    "    bash ./local/preprocess.sh ${conf_path} || exit -1\n",
    "fi\n",
    "if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then\n",
    "    # train model, all `ckpt` under `train_output_path/checkpoints/` dir\n",
    "    CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1\n",
    "fi\n",
    "if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then\n",
    "    # synthesize, vocoder is pwgan\n",
    "    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1\n",
    "fi\n",
    "if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then\n",
    "    # synthesize_e2e, vocoder is pwgan\n",
    "    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1\n",
    "fi\n",
    "if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then\n",
    "    # inference with static model\n",
    "    CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path} || exit -1\n",
    "fi\n",
    "```\n",
    "\n",
    "### 基于 CSMCS 数据集训练 Parallel WaveGAN 模型\n",
    "```bash\n",
    "git clone https://github.com/PaddlePaddle/PaddleSpeech.git\n",
    "cd examples/csmsc/voc1\n",
    "```\n",
    "根据 README.md, 下载 CSMCS 数据集和其对应的强制对齐文件, 并放置在对应的位置\n",
    "```bash\n",
    "./run.sh\n",
    "```\n",
    "`run.sh` 中包含预处理、训练、合成等步骤:\n",
    "```bash\n",
    "#!/bin/bash\n",
    "set -e\n",
    "source path.sh\n",
    "gpus=0,1\n",
    "stage=0\n",
    "stop_stage=100\n",
    "conf_path=conf/default.yaml\n",
    "train_output_path=exp/default\n",
    "ckpt_name=snapshot_iter_5000.pdz\n",
    "\n",
    "# with the following command, you can choice the stage range you want to run\n",
    "# such as `./run.sh --stage 0 --stop-stage 0`\n",
    "# this can not be mixed use with `$1`, `$2` ...\n",
    "source ${MAIN_ROOT}/utils/parse_options.sh || exit 1\n",
    "\n",
    "if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then\n",
    "    # prepare data\n",
    "    ./local/preprocess.sh ${conf_path} || exit -1\n",
    "fi\n",
    "if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then\n",
    "    # train model, all `ckpt` under `train_output_path/checkpoints/` dir\n",
    "    CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1\n",
    "fi\n",
    "if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then\n",
    "    # synthesize\n",
    "    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1\n",
    "fi\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# FAQ\n",
    "\n",
    "- 需要注意的问题\n",
    "- 经验与分享\n",
    "- 用户的其他问题"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 作业\n",
    "在 CSMSC 数据集上利用 FastSpeech2 和 Parallel WaveGAN 实现一个中文 TTS 系统。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 关注 PaddleSpeech\n",
    "请关注我们的 [Github Repo](https://github.com/PaddlePaddle/PaddleSpeech/)，非常欢迎加入以下微信群参与讨论：\n",
    "- 扫描二维码\n",
    "- 添加运营小姐姐微信\n",
    "- 通过后回复【语音】\n",
    "- 系统自动邀请加入技术群\n",
    "\n",
    "<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/bca0bc75dce14b53af44e374e64fc91aeeb13c075c894d6aabed033148f65377\" ></center>\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.4"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}