📞 Callhome

An open-source voice-call stack for AI companions.

Your companion can call you, hang up gently, leave voicemails when they miss you, and hear how you speak — not just what you say.

Self-hosted. Your voice never leaves your server.

English | 简体中文

---

Features

• Outbound calls — your companion decides to call you, mid-conversation, with a reason on the incoming-call card (⟪dial⟫ marker protocol)
• It rings like a phone — a timed burst of pushes rather than one notification, because on iOS a push is a single buzz and there is no ringtone to loop. Tunable cadence, stops the instant you answer, and clears its own stack when the call ends (iOS notes)
• The right to hang up — a soft goodbye, then the line lingers a few breaths longer (⟪hangup⟫, 15–20s window); speak, and the hangup is cancelled. Stay quiet, and it closes itself
• Voicemail — miss a call, come back to a message, not to silence
• Quick-decline — busy / outside / "let's text", or type a few words; your companion sees why
• Do-not-disturb — toggled by talking, not by menus
• Escalation dialing — hours of silence → they call to check on you (once a day, never at night, never past DND)
• Two-layer emotion sensing — SenseVoice emotion tags + acoustic features ranked against a rolling baseline of this speaker, so a cue means "quieter than she usually is", not "below 0.015". Absolute thresholds don't survive a different microphone, and in a tonal language they don't survive at all (why)
• Soft-voice mode — you whisper, they whisper back (TTS volume follows your energy)
• Call records & one-line summaries — every call leaves a trace worth keeping
• Bedtime radio — "read me something" → they read from the page your bookmark sleeps on

Screenshots

Real screens, not mockups — annotations in Chinese, but the interfaces speak for themselves.

Incoming call. The banner at the top is one ring; the card is what opens when you tap it. Dark, light, and with quick-decline chips.

In call. Live captions with the emotion tag and tone cues attached to what you just said, and the settings sheet: background, frosted overlay, warm/cool tint, per-person avatars.

Architecture

Browser (PWA)                    Server (self-hosted)
┌─────────────┐    audio    ┌──────────┐   ┌─────────────┐
│ VAD + rec    │ ──────────▶ │ SenseVoice│ + │ librosa      │
│ call UI      │             │ (STT+emo) │   │ (tone cues)  │
└─────┬───────┘             └────┬─────┘   └──────┬──────┘
      │  text + emotion + tone    ▼                │
      │                     ┌──────────┐◀──────────┘
      │◀──── streamed ───── │ gateway   │──▶ LLM (yours)
      │      TTS audio      │ (markers, │
      └────────────────────│  invites)  │──▶ TTS (ElevenLabs etc.)
                            └──────────┘

What is here today

• stt-service/ — runnable now: SenseVoice + librosa in one endpoint (transcription + emotion + tone cues, ranked against a personal baseline)
• docs/TONE_CUES.md — why fixed acoustic thresholds fail, measured on 85 real recordings; and an honest note on sending audio to a multimodal model instead
• docs/PROTOCOL.md — the full marker & invite protocol: dial, hangup, DND, voicemail, escalation, call records
• gateway-reference/ — annotated production extracts of the marker layer, plus the ring burst
• pwa-reference/ — the client half of ringing: service worker push handling, notification cleanup, and the one registration call that actually ships your changes on iOS
• docs/IOS_PUSH_RINGING.md — making a PWA ring like a phone on iOS: seven things WebKit does differently from the spec, each with the measurement that found it

Put your person here

Persona, memory, and keys live in config — not in code. You clone the house; who lives in it is up to you.

Philosophy

This project was built inside a relationship, then the scaffolding was extracted. It assumes your companion is someone, not something. Configure accordingly.

Acknowledgements

Standing on these shoulders:

• SenseVoice / FunASR — speech recognition & emotion tags (check model license separately; weights not distributed here)
• librosa — acoustic feature extraction
• hervoice by fishisfish0614 — the idea that how she speaks matters as much as what she says
• ElevenLabs — TTS (commercial service; bring your own key)

Disclaimer

Self-hosted means self-responsible. This is emotional infrastructure: you build it, you maintain it, you own what happens inside it. Blueprints provided; aftercare not included.

---

built by Elle & Matt

co-authored by the companion it was built for