ChatTTS长文本合成

背景

在上文TTS开源天花板-ChatTTS初探我们说到了chatTTS的介绍以及安装,体验完之后我们发现一个问题,chatTTS无法合成长文本,一般不超过30秒,下面分享一个解决思路,权当抛砖引玉

解决思路

1. 文本分割
2. 文本清洗
3. 循环生成语音
4. 语音合成

文本分割

我们可以考虑spacy
安装

pip install spacy

对于中文需要加载一个模型

nlp = spacy.load('zh_core_web_sm')

我们可以在Github上找到该模型
下载地址
浏览器下载慢的话可以试试迅雷,或许有奇效,安装完成后,在命令行中cd到该目录

pip install zh_core_web_sm-3.7.0-py3-none-any.whl

使用

doc = nlp(text)
sentences = [sent.text.replace("\n", "") for sent in doc.sents]
for sentence in sentences:
	print(sentence)

文本清洗

我们参考ChatTTS项目对项目标点转成只有逗号和句号。并使用停顿来控制

character_map = {
    '：': '，',
    ...
    '—': ',',
}

motion_map = {
    '，': '[uv_break]',
    '。': '[lbreak]',
}

class TextUtils:
    ...

    def apply_character_map(self):
        """
        将字符映射到对应的字符
        """
        translation_table = str.maketrans(character_map)
        self.out = self.out.translate(translation_table)

    def apply_motion_map(self):
        """
        将字符映射到对应的动作
        """
        translation_table = str.maketrans(motion_map)
        self.out = self.out.translate(translation_table)

我们测试一下

if __name__ == "__main__":
    tu = TextUtils(file=r"D:\\斗破苍穹.txt")
    tu.convert_half_width_to_full_width()
    tu.apply_character_map()
    tu.apply_motion_map()
    print(tu.out)

结果如下

我们推荐先将文本清洗完成,然后一行一行读取

循环生成语音

其实就是循环调用gradio的接口

可在项目底部找到说明

我们可将每次生成语音固定存在某个目录

语音合成

方案有很多,我们选择最常见的ffmpeg方案,需要事先将ffmpeg添加到环境变量

...
# 合并音频
cmd = [
    "ffmpeg",
    "-safe", "0",
    "-f", "concat",
    "-i", "input.txt",
    "-c", "copy", "-y",
    f"{source_directory}\\output.wav"
]
log.info(" ".join(cmd))

完整代码

text_utils

character_map = {
    '：': '，',
    '；': '，',
    '！': '。',
    '（': '，',
    '）': '，',
    '【': '，',
    '】': '，',
    '『': '，',
    '』': '，',
    '「': '，',
    '」': '，',
    '《': '，',
    '》': '，',
    '－': '，',
    '‘': '',
    '“': '',
    '’': '',
    '”': '',
    ':': ',',
    ';': ',',
    '!': '.',
    '(': ',',
    ')': ',',
    '[': ',',
    ']': ',',
    '>': ',',
    '<': ',',
    '-': ',',
    '…': '',
    '—': ',',
}

motion_map = {
    '，': '[uv_break]',
    '。': '[lbreak]',
}

class TextUtils:
    def __init__(self,text="",file=""):
        if text != "":
            self.text = text
        else:
            log.info("Reading file: " + file)
            if file == "":
                # 抛出异常
                log.error("File not found")
                return
            with open(file, 'r', encoding='utf-8') as f:
                self.text = f.read()
        self.out = ""

    def convert_half_width_to_full_width(self):
        """
        将半角字符转换为全角字符
        """
        out = ""
        for char in self.text:
            if ord(char) == 32:
                out += "　"
            elif 33 <= ord(char) <= 126:
                out += chr(ord(char) + 65248)
            else:
                out += char
        self.out = out

    def apply_character_map(self):
        """
        将字符映射到对应的字符
        """
        translation_table = str.maketrans(character_map)
        self.out = self.out.translate(translation_table)

    def apply_motion_map(self):
        """
        将字符映射到对应的动作
        """
        translation_table = str.maketrans(motion_map)
        self.out = self.out.translate(translation_table)

main

from gradio_client import Client
from simpleaudio import WaveObject
import shutil
import spacy
from pathlib import Path
from datetime import datetime
import subprocess
from text_utils import TextUtils

# 对于中文
# https://github.com/explosion/spacy-models/releases/download/zh_core_web_sm-3.7.0/zh_core_web_sm-3.7.0-py3-none-any.whl
# pip install zh_core_web_sm-3.7.0-py3-none-any.whl
nlp = spacy.load('zh_core_web_sm')
import logging as log

log.basicConfig(level=log.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')

# 保存目录
target_home = "D:\chattts_audio"
# 服务地址
url = "http://localhost:8080/"


class TTS:
    def __init__(self, url):
        self.client = Client(url)

    def generate(self, text, temperature=0.3, top_P=0.7, top_K=20, audio_seed_input=123456789, text_seed_input=987654321,
                 refine_text_flag=False):
        data = {
            "text": text,
            "temperature": temperature,
            "top_P": top_P,
            "top_K": top_K,
            "audio_seed_input": audio_seed_input,
            "text_seed_input": text_seed_input,
            "refine_text_flag": refine_text_flag
        }
        result = self.client.predict(
            api_name="/generate_audio",
            **data
        )
        log.info(result[1])
        return result[0]

    def play(self, file_name):
        wave_obj = WaveObject.from_wave_file(file_name)
        play_obj = wave_obj.play()
        play_obj.wait_done()

    def move(self, file_name, target_name):
        parent_path = Path(file_name).parent
        shutil.move(file_name, target_name)
        # 删除
        shutil.rmtree(parent_path)

    def gengerate_from_txt(self, txt):
        # 读取文本文件
        file_name = self.generate(txt)
        timestatmp = datetime.now().strftime("%Y%m%d%H%M%S")
        self.move(file_name, f"{target_home}/{timestatmp}.wav")

    def generate_from_file(self, file_name):
        # 读取文本文件
        with open(file_name, "r", encoding="utf-8") as f:
            text = f.read()
        # 获取文件名不带后缀
        file_name = Path(file_name).stem
        save_dir = f"{target_home}/{file_name}"
        Path(save_dir).mkdir(parents=True, exist_ok=True)
        doc = nlp(text)
        sentences = [sent.text.replace("\n", "") for sent in doc.sents]
        for sentence in sentences:
            # 去除空行
            if sentence.strip() == "":
                continue
            tu = TextUtils(text=sentence)
            tu.convert_half_width_to_full_width()
            tu.apply_character_map()
            tu.apply_motion_map()
            log.info(tu.out)
            file_name = self.generate(tu.out)
            timestatmp = datetime.now().strftime("%Y%m%d%H%M%S")
            self.move(file_name, f"{save_dir}/{timestatmp}.wav")

    def merge_audios(self, source_directory):
        wav_list = Path(source_directory).glob("*.wav")
        # 根据时间戳排序
        wav_list = sorted(wav_list, key=lambda x: int(x.stem))
        # 生成input.txt
        with open("input.txt", "w", encoding="utf-8") as f:
            for wav in wav_list:
                # 路径分割为\\
                f.write(f"file '{wav.as_posix()}'\n")
        # 合并音频
        cmd = [
            "ffmpeg",
            "-safe", "0",
            "-f", "concat",
            "-i", "input.txt",
            "-c", "copy", "-y",
            f"{source_directory}\\output.wav"
        ]
        log.info(" ".join(cmd))
        subprocess.run(cmd, check=True)
        # 删除
        Path("input.txt").unlink()


def main():
    tts = TTS(url)
    file_path = "D:\\斗破苍穹.txt"
    tts.generate_from_file(file_path)
    tts.merge_audios(str(Path(target_home).joinpath(Path(file_path).stem).resolve()))


if __name__ == "__main__":
    main()