tts-mcp-server
Mentioned in the project roadmap (P2) as the planned compositing backend for a video translation-and-dubbing workflow, where synthesized speech and generated subtitles would be combined into finished videos via ffmpeg.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@tts-mcp-server用会议预设把这段文本转成带字幕的语音:各位同事早上好,本次周会将讨论三个议题。"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
tts-mcp-server — 中文TTS聚合MCP Server
一个 MCP Server,把中文语音合成 + 精确字幕对齐做成 Agent 可调用的工具。 多引擎聚合:Edge-TTS(在线免费) + Piper(离线免费),统一音色目录,用户无感切换。
演示

文本 → 语音 + 逐词对齐字幕(真实引擎时间戳,非估算)
Related MCP server: GPT-SoVITS MCP Server
为什么做这个
现有 TTS MCP 生态里:
Edge-TTS 的 MCP 都是单引擎薄壳(最高 8★、只活 2 天、只有 1 个
text_to_speech工具)没有任何一个能生成字幕 —— 但字幕对齐才是视频/播客/有声书的刚需
CosyVoice 等本地克隆的中文 MCP 有效实现 = 0
本项目的杀手锏:synthesize_with_subtitles —— 文本转语音的同时,从引擎的真实边界事件生成逐词对齐的 SRT 字幕,可直接烧录到视频。
核心能力
工具 | 说明 |
| 列出全部音色(14个中文,含方言/港台),按引擎/性别/方言过滤 |
| 文本 → MP3 + SRT,多引擎聚合,支持预设 |
| 文本 → 语音 + 引擎级时间戳字幕(Edge用WordBoundary词级,Piper用音素对齐字符级) |
| 批量并行合成:多段文本并发TTS(实测提速13倍),每段独立音频+字幕 |
Resource: | 音色目录 + 5个预设(新闻/小说/方言/会议/离线) |
字幕对齐原理
不是估算,是引擎的真实时间戳:
Edge-TTS: WordBoundary 事件 → 词级 {start_s, end_s}
Piper: 本地离线,字符级时间戳(无网可用,CPU极快 RTF<0.1)
→ 统一合并为 SRT 行(硬断点优先 + 字符/时长双限)实测:最后一条字幕的结束时间 = 音频真实总时长,误差 <0.5s。
安装
pip install tts-mcp-server # 干净底座(字幕算法 + 框架)
pip install tts-mcp-server[edge] # + Edge-TTS 引擎(推荐,默认)
pip install tts-mcp-server[piper] # + Piper 离线引擎(可选,断网兜底)使用
Claude Desktop
claude_desktop_config.json:
{
"mcpServers": {
"tts-server": {
"command": "tts-mcp",
"args": [],
"env": { "TTS_OUTPUT_DIR": "/path/to/tts_out" }
}
}
}Cursor
.cursor/mcp.json:
{
"mcpServers": {
"tts-server": {
"type": "stdio",
"command": "tts-mcp",
"args": []
}
}
}演示用例:会议纪要 → 带字幕的播客
调用 synthesize_with_subtitles(
text="各位同事早上好,本次周会将讨论三个议题……",
preset="meeting"
)
→ 得到 meeting.mp3 + meeting.srt
→ 播放器加载 SRT,逐句跟读,重点可跳转预设
预设 | 音色 | 语速 | 场景 |
news | 云扬(男) | +5% | 新闻播报 |
novel | 云希(男) | 0 | 有声小说 |
dialect | 晓北(东北话) | 0 | 方言演示 |
meeting | 晓晓(女) | +5% | 会议纪要 |
offline | 华炎(Piper) | 0 | 断网离线 |
路线图
P0: Edge-TTS 引擎 + 字幕对齐 + 预设
P1: 百度 TTS 接入(个人实名免费额度)+ whisper 转写
P2: 视频翻译配音工作流(ffmpeg 合成)
License
MIT。引擎适配层(engines/edge 依赖 edge-tts GPL-3.0、engines/piper 依赖 piper-tts GPL-3.0)通过 optional-dependencies + 晚绑定隔离,主包许可证干净。
系列(Chinese MCP Suite)
中文内容创作/工具 MCP 全家桶,全本地、CPU、离线:
tts-mcp-server(本仓库)— TTS聚合+字幕对齐
idphoto-mcp — 本地证件照
audio-post-mcp — 音频后期
video-workflow-mcp — 视频工作流
poetry-mcp — 古诗文
script-mcp — 口播文案
asset-mcp — 素材库
idiom-mcp — 成语典故
weather-mcp — 天气/空气
This server cannot be deployed
Maintenance
Related MCP Connectors
Generate English and Chinese speech, split long text, track batch jobs, and download audio.
Use VoiceCheap’s media tools from your AI agent: transcribe audio, video and YouTube; create, translate and add subtitles to videos; generate YouTube chapters; create speech, design voices or clone voices with permission; generate images and thumbnails; translate and dub videos with optional lip sync; and publish completed projects to your connected channels. Manage project status, transcripts, glossaries and workspace settings through the tools you authorize. Available tools depend on your OAuth permissions and account access. Generation uses existing credits; purchases take place on the VoiceCheap website.
Transcribe, subtitle, dub and lip-sync media; create speech, voices, images and thumbnails.
AI Voice Production Agent: picks from 30+ TTS models, fixes pronunciation, validates every line.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to convert text to high-quality speech audio using MeloTTS. Automatically splits long texts into segments, generates WAV files, and merges them using ffmpeg with support for multiple languages and customizable speech parameters.MIT
- FlicenseNot gradedqualityDmaintenanceEnables text-to-speech synthesis and voice cloning through GPT-SoVITS API integration. Supports multiple languages (Chinese, English, Japanese, Korean, Cantonese), dynamic model switching, and reference audio-based voice quality replication.4-
- FlicenseNot gradedqualityDmaintenanceEnables voice cloning and speech synthesis through Alibaba Cloud's Bailian and DashScope platforms. It provides tools to create, manage, and synthesize audio using custom cloned voice profiles.-
- AlicenseNot gradedqualityDmaintenanceEnables speech-to-text and text-to-speech conversion using OpenAI-compatible APIs. Supports customizable models, voices, and output directories.GPL 3.0