chatts长文本测试代码。

import ChatTTS
import torch
import numpy as np
import torchaudio
import re
import time
# 记录开始时间
start_time = time.time()

# 设置环境变量以避免内存碎片化
import os
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'

# 使用 CPU 进行计算
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

chat = ChatTTS.Chat()
chat.load(compile=False)  # Set to True for better performance

text = """
小品文的体裁多种多样，有游记、速写、随笔、杂感（或称杂文），寓言、序、尺牍等。这些文体在古人看来，与诗文正统文学相比是不登大雅之堂的，故在古代分类的文体论中，不见有单独的“小品文”一类。中国的文体论，发轫于魏晋，盛于齐梁以后。曹丕的《典论·论文》中把文分为奏议、书论、铭诔；陆机的《文赋》将文体分为九类，即赋、碑、诔，铭、箴、颂、论、奏、说，属于“文”的有六种；晋·挚虞的《文章流别论》中也把文分为九类：诗、颂、赋、七、箴、铭、诔、哀、碑，属于文的有五种。至南北朝，刘勰在《文心雕龙》中提出了著名的文、笔说，“无韵者笔也”，笔指各种散体文，他所列的二十种文体中，属于无韵文的有十种，还出现了“杂文”这一名称（不过它属于韵文），就是还没有“小品文”。梁·昭明太子的《文选》将诗文分为三十八类，属于散文的有二十一种；到了明代吴讷的《文章辨体》和徐师曾的《文体明辨》，入选的文体更广泛，分类更细密，前者共分文体五十九类，后者搜罗文体一百二十七类,还是没有提及“小品文”这一名称。 [2]
"""
# 使用固定音色
guding_spk = torch.load("speaker/seed_2279.pt", map_location=device, weights_only=True)

# 相关参数，见文档说明
params_infer_code = ChatTTS.Chat.InferCodeParams(
    spk_emb=guding_spk,  # add sampled speaker
    temperature=.3,     # using custom temperature
    top_P=0.7,          # top P decode
    top_K=20,           # top K decode,

)

# 根据效果自行选择
# use oral_(0-9), laugh_(0-2), break_(0-7)
# to generate special token in text to synthesize.
params_refine_text = ChatTTS.Chat.RefineTextParams(
    prompt='[oral_0][laugh_0][break_6]',
)
def remove_bracketed_content(text):
    """
    移除文本中所有以 [ 开头和 ] 结尾且中间包含字母或数字的内容。
    """
    return re.sub(r'\[\w+\]', '', text)

def replace_punctuation_with_uv_break(text):
    """
    将文本中的标点符号替换为 [uv_break]。
    """
    # 定义常见的标点符号
    punctuation_pattern = r'[（）；;·：“”【】《》、…：—]'
    return re.sub(punctuation_pattern, ',', text)
text1=remove_bracketed_content(text)
# print(text1)
text=replace_punctuation_with_uv_break(text1)
# print(text)
def text_to_sentences(text, max_length=180, sub_max_length=90):
    # 去除多余的空行和缩进
    text = re.sub(r'\s+', ' ', text).strip()

    # 使用正则表达式根据标点符号分割句子
    sentence_endings = re.compile(r'(?<=[。！？])')
    raw_sentences = sentence_endings.split(text)

    sentences = []
    current_sentence = ""

    for sentence in raw_sentences:
        sentence = sentence.strip()
        if not sentence:
            continue

        # 如果当前句子加上新句子超过 sub_max_length，则先保存当前句子
        if len(current_sentence) + len(sentence) + 1 > sub_max_length:
            if current_sentence:
                sentences.append(current_sentence)
                current_sentence = ""

        # 如果新句子本身超过 sub_max_length，则进一步分割
        if len(sentence) > sub_max_length:
            sub_start = 0
            while sub_start < len(sentence):
                sub_end = sub_start + sub_max_length
                if sub_end < len(sentence):
                    sub_end = sentence.rfind(' ', sub_start, sub_end)
                    if sub_end == -1:
                        sub_end = sub_start + sub_max_length
                sub_sentence = sentence[sub_start:sub_end].strip()
                if sub_sentence:
                    sentences.append(sub_sentence)
                sub_start = sub_end + 1
        else:
            # 如果当前句子为空，则直接赋值
            if not current_sentence:
                current_sentence = sentence
            else:
                current_sentence += " " + sentence

    # 添加最后一个句子
    if current_sentence:
        sentences.append(current_sentence)

    return sentences
# def text_to_sentences(text, max_length=180, sub_max_length=90):
#     # 去除多余的空行和缩进
#     text = re.sub(r'\s+', ' ', text).strip()
#
#     sentences = []
#     start = 0
#     while start < len(text):
#         end = start + max_length
#         if end < len(text):
#             # 找到最后一个空格的位置，确保不会截断单词
#             end = text.rfind(' ', start, end)
#             if end == -1:
#                 end = start + max_length
#         sentence = text[start:end].strip()
#         if len(sentence) > sub_max_length:
#             # 进一步分割长句子
#             sub_start = 0
#             while sub_start < len(sentence):
#                 sub_end = sub_start + sub_max_length
#                 if sub_end < len(sentence):
#                     sub_end = sentence.rfind(' ', sub_start, sub_end)
#                     if sub_end == -1:
#                         sub_end = sub_start + sub_max_length
#                 sub_sentence = sentence[sub_start:sub_end].strip()
#                 if sub_sentence:
#                     sentence = sentence
#                     sentences.append(sub_sentence)
#                 sub_start = sub_end + 1
#         else:
#             if sentence:
#                 sentence=sentence
#                 sentences.append(sentence)
#         start = end + 1
#
#     return sentences

# def text_to_sentences(text):
#     # 去除多余的空行和缩进
#     text = re.sub(r'\s+', ' ', text).strip()
#
#     # 使用中文标点符号分割文本，并去除空字符串
#     punctuation_pattern = r'[，。！？；：“”‘’（）【】《》、…：—]'
#     sentences = [sentence.strip() for sentence in re.split(punctuation_pattern, text) if sentence.strip()]
#
#     return sentences


# 将文本分割成列表
texts = text_to_sentences(text)
print(texts)
wavs = []
for text in texts:
    wav = chat.infer(text, skip_refine_text=True, params_refine_text=params_refine_text, params_infer_code=params_infer_code)
    # 确保音频数据是一维数组
    if wav.ndim > 1:
        wav = wav.squeeze()
    wavs.append(wav)
def merge_wavs(wavs, output_file, sample_rate=24000):
    # 将所有wav数据连接成一个numpy数组
    merged_wav = np.concatenate(wavs, axis=0)

    # 转换为torch张量并添加通道维度
    merged_wav_tensor = torch.from_numpy(merged_wav).unsqueeze(0)

    # 保存合并后的音频文件
    torchaudio.save(output_file, merged_wav_tensor, sample_rate)

# def merge_wavs(wavs, output_file, sample_rate=24000):
#     # 检查所有音频数据的形状
#     max_length = max(wav.shape[0] for wav in wavs)
#     standardized_wavs = []
#     for wav in wavs:
#         if wav.shape[0] < max_length:
#             # 填充较短的音频
#             padded_wav = np.pad(wav, (0, max_length - wav.shape[0]), mode='constant')
#             standardized_wavs.append(padded_wav)
#         else:
#             standardized_wavs.append(wav)
#
#     # 将所有wav数据连接成一个numpy数组
#     merged_wav = np.concatenate(standardized_wavs, axis=0)
#
#     # 转换为torch张量并添加通道维度
#     merged_wav_tensor = torch.from_numpy(merged_wav).unsqueeze(0)
#
#     # 保存合并后的音频文件
#     torchaudio.save(output_file, merged_wav_tensor, sample_rate)

try:
    merge_wavs(wavs, "荷塘月色_children.wav", 24000)
except Exception as e:
    print(f"Error merging wavs: {e}")
finally:
    # 释放未使用的缓存
    torch.cuda.empty_cache()


# 记录结束时间
end_time = time.time()

# 计算并打印耗时
elapsed_time = end_time - start_time
print(f"Total time taken: {elapsed_time:.2f} seconds")