This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.

voice

by BioMediChen

Overview Schema Source Related Servers Discussions Admin

Python

Local

OpenCode 语音控制（离线 ASR）

license platform python

按住 F8 或鼠标侧键说话 → 松开 → SenseVoice 本地转写 → 文本自动作为用户消息投递到当前 OpenCode 会话。

ASR：sherpa-onnx + SenseVoice-small (int8)，纯 CPU，中文/中英混说，本机实测 11s 音频转写 0.28s
交互：全局热键 / 鼠标侧键按住说话（push-to-talk），无需唤醒词
投递：OpenCode HTTP API（Basic auth，自动从 opencode service status 发现端口）
打断：说“停止 / 取消”等词直接中断当前 agent 回合，不投递文本
纠错：config.toml 里的 [terms] 把常见误识别替换成代码写法（如“use state”→useState）
MCP：voice 服务器提供 status / listen / interrupt，listen 让 agent 等待用户口头确认（该次语音不会自动提交）

src/voice_asr/
  asr.py            SenseVoice 封装
  audio.py          麦克风采集、静音裁剪、重采样
  corrections.py    术语纠错
  opencode_client.py OpenCode 会话发现与 prompt 投递
  daemon.py         热键/鼠标守护进程 + 本地控制接口 (127.0.0.1:8756)
  mcp_server.py     MCP stdio 服务器（工具层）
  check.py          自检：设备列表 / 模型 / 转写
config.toml         配置
models/sensevoice/  SenseVoice int8 模型 (239MB) + tokens
logs/               运行日志
scripts/            MCP/投递自检、开机自启安装脚本

快速开始（从零克隆）

git clone https://github.com/BioMediChen/opencode-voice.git E:\voice
cd E:\voice

# 1) Python 环境（uv 会自动装 3.12）
uv sync

# 2) 下载 SenseVoice int8 模型（约 239MB，国内可用 hf-mirror）
New-Item -ItemType Directory -Force models\sensevoice | Out-Null
curl.exe -L -o models\sensevoice\model.int8.onnx https://hf-mirror.com/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/model.int8.onnx
curl.exe -L -o models\sensevoice\tokens.txt https://hf-mirror.com/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/main/tokens.txt

# 3) 自检
uv run voice-check --record 4
uv run python scripts\selftest_e2e.py

# 4) 启动守护进程
uv run voice-daemon

平台：Windows 10/11（音频提示音、全局热键依赖 Windows）。Linux 下需自行处理 winsound 和音频设备。

实测效果

[recording:mouse:x2] ... release to transcribe
[asr] 1.38s  测试us stay和优 effect.
[terms] ... -> 测试useState和useEffect.
[sent -> ses_e6047cd3...] opencode MCP 语音转文字接入

SenseVoice-small int8 在 Ryzen 7 8840U（8C/16T）上的实测：5.7 秒语音转写 0.14 秒， RTF ≈ 0.03（约为实时的 30 倍），全程离线、无 GPU。

作为命令行工具安装（可选）

不想克隆整个仓库时：

uv tool install git+https://github.com/BioMediChen/opencode-voice

# 配置和模型放在固定目录（模型路径相对 config.toml 所在目录解析）
$env:VOICE_CONFIG="E:\voice\config.toml"
voice-daemon

安装后可用：voice-daemon、voice-mcp、voice-check。 MCP 注册时把 command 换成本机的 voice-mcp 可执行文件即可。

使用

# 自检：列出输入设备，用麦克风录 4 秒并转写
uv run voice-check --record 4

# 转写一个 wav 文件
uv run voice-check --wav whisper.cpp\samples\jfk.wav

# 端到端自检（不消耗模型额度）
uv run python scripts\selftest_e2e.py

# 前台启动守护进程（调试用，能看实时日志）
uv run voice-daemon

# 静默启动（日常使用）
start-voice-silent.cmd

启动后按住 F8 或鼠标侧键说话，松开即发送（开始/结束各有一声提示音）。

# 查看状态
curl.exe http://127.0.0.1:8756/status

# 无麦克风时直接测试投递
curl.exe -X POST http://127.0.0.1:8756/submit -H "Content-Type: application/json" -d "{\"text\":\"你好\"}"

配置要点（config.toml）

项	说明
`hotkey.key`	键盘触发键，`f1`..`f12` / `caps_lock` / `scroll_lock` / 单字符
`hotkey.mouse`	鼠标触发键，`x1`(后) / `x2`(前) / `middle`，`[]` 关闭
`audio.device`	留空用系统默认；填序号或名称片段（如 `HECATE`）
`commands.stop`	说完即打断的短语（整句匹配，忽略标点）
`terms`	误识别 → 目标写法；英文规则按单词边界匹配，不会破坏 `useState`
`opencode.session_mode`	`auto`（跟随活跃/最近会话，支持多项目）/ `directory`（限定目录）/ `pinned`（固定 `session_id`）
`opencode.directory`	`directory` 模式下的目标目录
`opencode.session_id`	`pinned` 模式下的固定会话
`asr.language`	`auto` / `zh` / `en` / `ja` / `ko` / `yue`

修改配置后重启守护进程。注意鼠标侧键同时也会被浏览器/资源管理器当作“前进/后退”，如果觉得干扰，可以把 hotkey.mouse 改成 ["middle"] 或 []。

自检与术语维护

# 端到端自检：守护进程 / 会话创建 / 投递 / 纠错 / 停止分支（不消耗模型额度）
uv run python scripts\selftest_e2e.py

# MCP 服务器工具列表自检
uv run python scripts\selftest_mcp.py

# 扫描日志，找出还没被 [terms] 覆盖的英文词，给出待补建议
uv run python scripts\suggest_terms.py

[terms] 的格式是「正确写法 = [可能被听成的说法]」，以 use 开头的两词名称会自动展开 use/us/u/you/优/优思/优斯 与 state/stay/stat、effect/affect 等音近组合，所以通常只需写标准说法，例如：

[terms]
"useState" = ["use state"]
"OpenCode" = ["open code", "opencode"]
"E:\\voice" = ["e 盘 voice", "e voice"]

多项目切换

opencode.session_mode = "auto" 时的选择顺序：

配置了 session_id → 固定投递到该会话
当前有且仅有一个会话在运行 → 投递到它（语音打断/追加指令就靠这条）
否则投递到最近更新的会话（跨项目生效，跟着你正在用的会话走）

如果想锁定某个项目，改成 "directory" 并设置 opencode.directory；想锁定某个会话，改成 "pinned" 并填写 opencode.session_id。

MCP 注册

全局配置 ~/.config/opencode/opencode.json 已注册：

{
  "mcp": {
    "servers": {
      "voice": {
        "type": "local",
        "command": ["E:\\voice\\.venv\\Scripts\\python.exe", "-m", "voice_asr.mcp_server"],
        "cwd": "E:\\voice",
        "environment": { "PYTHONUTF8": "1", "PYTHONIOENCODING": "utf-8" }
      }
    }
  }
}

MCP 服务器在需要时会自动拉起守护进程（VOICE_NO_AUTOSTART=1 可关闭）。

开机自启

已通过 scripts/install-autostart.ps1 在启动文件夹创建 OpenCode Voice.lnk。重新安装/更换路径时重跑：

powershell -ExecutionPolicy Bypass -File scripts\install-autostart.ps1

许可证

MIT

This server cannot be deployed

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

Maintenance

ActivityMaintained

ResponsivenessNo issues

Resources

Related MCP Connectors

mosADD — the comms layer for AI agents B
com.mosadd
Command your AI agents by voice: PTT rooms, channels, direct messages, agent email, memory (mRAG).
Communication Speech Processing Email
10311
https://mcp.mosadd.com/mcp
Speechtotext Mcp
io.github.varvararatta
MCP server for Speech-to-Text
Speech Processing AI & Machine Learning
2
https://speechtotext_mcp.mcpize.run/mcp
Superforms B
co.superforms
Human-input bridge for AI agents with voice-first answer links, MCP tools, and HTTP APIs.
Agent Orchestration Speech Processing App Automation
112
https://superforms.co/mcp
Heard A
dev.heard
Let cloud agents talk to you: progress updates spoken on your Mac, and replies you send back.
Developer Tools
31
https://api.heard.dev/v1/mcp/agent

Related MCP Servers

EarSay MCP
AlastorMordrek
A
license
A
quality
D
maintenance
Provides voice transcription control and polling for MCP-compatible agents, enabling start/stop/pause/resume and retrieval of new text via tools.
Updated 3 months ago (2026-07-21 20:57 UTC)
8
MIT
genpark-conversational-bargeofficial
Audio Processing Speech Processing Autonomous Agents
Alpha-Park
F
license
Not graded
quality
A
maintenance
Enables real-time full-duplex voice agents to classify user interruptions, backchannels, and acoustic echo, and to arbitrate turn-taking with context rollback via MCP.
Updated a month ago (2026-09-28 10:15 UTC)
7
-
stts-mcp
Speech Processing Text-to-Speech
markkennethbadilla
A
license
Not graded
quality
B
maintenance
Enables coding agents to hold spoken conversations by exposing speech-to-text and text-to-speech tools over MCP, routing them to a local daemon that drives a Chrome window using on-device speech recognition and Piper synthesis. It supports listening, speaking, and reading files aloud within a Claude Code plugin or an MCP gateway.
Updated 4 days ago (2026-10-30 22:50 UTC)
2 npm
MIT
Utterlane Live
Speech Processing Text-to-Speech Audio Processing
tuncertimur
A
license
Not graded
quality
A
maintenance
Exposes a local, hands-free voice room to any MCP client, letting agents read speaker-labelled conversation turns, wait for the next turn, list who is present, and speak aloud through local text-to-speech. Speech recognition, speaker identification and voice output all run on the user's own machine, so agents receive text with speaker names rather than audio.
Updated 15 days ago (2026-10-19 21:26 UTC)
Apache 2.0