终于有人把“本地大模型打真实电话”做出来了，而且完全开源！

以往体验市面上的各类 AI 电话助手，最让人顾虑的就是隐私安全：通话录音、转写文本、沟通对象全都会留在第三方云端。开发者刚刚开源的 ai-phone 则彻底打通了本地闭环——让你的 Mac 直接接管真实电话线路，所有推理均在本地完成，通话数据完全不出机器。

🛠️ 核心技术栈（基于 Apple Silicon & MLX）：

语音识别（STT）： Parakeet TDT 0.6B 本地运行，极速语音转文字

语言模型（LLM）： 默认本地运行 Qwen3.5 4B（32GB 内存支持 9B 4-bit），同时也支持接入自己的 Claude 或 ChatGPT 端点

语音合成（TTS）： Kokoro 82M / Phonon，生成自然流畅的声音

通信接入： 通过 SIP 协议直连 Telnyx 或 Twilio，直接购买或绑定真实电话号码（月租约 1 美元）

✨ 核心亮点：

低延迟与实时打断： 在 Apple Silicon 芯片上实测首字响应延迟约 1~1.5 秒；支持实时插话打断（Interruption），一旦你开口，AI 会立刻停下倾听。

支持外部 Agent 联动： 提供了 call、listen、say、hangup 等标准化 CLI 接口，可以直接让 Claude Code、Codex 等编码智能体通过命令编排通话，自动完成预约餐厅、电话咨询等任务。

本地守护进程： 提供向导式 CLI 配置，一行命令拉起本地守护进程，静默接听来电或随时发起外呼。

⚠️ 当前要求与限制：

仅支持搭载 Apple Silicon 的 macOS 系统

目前为单路通话，优先支持英文环境

项目刚刚上线，对本地 AI 语音助手与 SIP 通信落地感兴趣的朋友值得关注：https://github.com/oxnw/ai-phone

oxnw/ai-phone

Source: https://github.com/oxnw/ai-phone

ai-phone

An AI phone agent that runs on your computer.

Documentation · Architecture · AGPL-3.0

Installation

git clone https://github.com/oxnw/ai-phone.git
cd ai-phone
./scripts/install.sh
uv run ai-phone setup

Setup lets you choose incoming calls, outgoing calls, or both, then configures the model, line and voice. It recommends a local model based on detected RAM and downloads the required weights. Use uv run ai-phone setup --plain for plain prompts.

The current speech backends use MLX on Apple Silicon, even when a cloud model handles the conversation. Supporting Linux or Windows would require different speech backends and changes to installation and platform integration. The platform requirement applies to the computer running ai-phone; callers can use any phone or SIP client.

Usage

Start the daemon after setup:

uv run ai-phone run

Keep it running while you want to receive calls. For outgoing calls only, start it when needed and stop it with Ctrl-C afterward. Setup does not install a background service.

From another terminal in the repo, give the phone agent instructions for a call:

uv run ai-phone call +15555550100 --json \
  --prompt "Ask whether they have a table for two at eight tonight."

Replace the example number with the number you want to call. The command returns the transcript when the call ends. Every outgoing call starts with an AI disclosure. Incoming calls use the greeting and instructions you set during setup. English and one call at a time are currently supported.

For a free local test, enable incoming calls, select A SIP address in setup, and keep the name agent:

pjsua --local-port 5070 sip:[email protected]

Run setup again to change the voice, model, call directions or caller ID, then restart the daemon.

Agent integration

An agent with shell access can use relay mode to control each turn:

uv run ai-phone call +15555550100 --relay --json
uv run ai-phone say "Hi, I'd like to book a table for two." --json
uv run ai-phone listen --json
uv run ai-phone say "Eight o'clock, please." --json
uv run ai-phone hangup --say "Thank you. Goodbye." --json

With no active call, listen --json waits for an incoming call and handles it in relay mode. Each command with --json returns one JSON object on stdout; call also sends progress events to stderr. An unanswered relay turn times out after 25 seconds by default. See the command reference for call IDs, transcript retrieval, timeouts and exit codes.

Models

Speech recognition and synthesis stay local. A remote language model receives the call’s text, instructions and conversation history. An external agent’s cloud provider also sees any transcripts passed to it. Phone audio travels over SIP/RTP; carrier charges and subscription limits apply.

Documentation

• Usage guide: lines, forwarding, caller ID, models and CLI reference.
• Example configuration: available settings.
• Architecture: components and call flow.
• NAT checklist: network configuration.
• Measurements: latest local call and interruption results.
• Contributing

Install and setup have been checked from a fresh clone. Local call and interruption tests have run; live carrier and end-to-end agent acceptance checks remain pending. The latest local run measured 1.50 s median first-audible response under heavy memory pressure; it missed the one-second target. Full conditions and earlier results are documented.

License

AGPL-3.0-or-later. The bundled PJSIP interface files use GPLv2-or-later; see the license notes.