message.comDevelopers
Developers/AI/Voice cloning

AI voice cloning Planned

Clone your brand voice from a fifteen-second audio sample. The cloned voice plays on Smart Hold, AI receptionist, voicemail greetings, and TTS confirmations on outbound calls. Powered by our neural voice synthesis pipeline; ships as a free feature.

Voice quality is a massive differentiator on phones. Intercom Fin uses generic stock voices. We let every brand sound like itself. Choosing the right synthesis provider is what makes this free at our price points.

Why this is free

Voice cloning costs us pennies per minute at our synthesis provider's wholesale rates. Bundling it removes the discount-vs-premium friction that competitors charge for. Brands that sound like themselves convert better; we'd rather you ship.

Recording the sample

Fifteen seconds of clear speech, recorded in a quiet room, single speaker. The voice should read sample copy that covers a range of phonemes; we provide a script in the dashboard. Avoid music, background noise, and reverb. Studio microphones are ideal but a modern smartphone in a quiet room works well enough.

  • Minimum 10 seconds, recommended 15 to 30 seconds.
  • Single speaker only.
  • WAV (16 kHz or higher) or MP3 (192 kbps or higher).
  • Up to 5MB.

API

Most teams upload through the dashboard, but the REST endpoint is available for power users (for example, an internal tool that lets each location at a franchise upload their own voice).

POST /api/v1/voice-profiles
curl -X POST 'https://app.message.com/api/v1/voice-profiles' \
  -H 'Authorization: Bearer YOUR_WORKSPACE_JWT' \
  -F 'displayName=Acme Brand Voice' \
  -F 'personalityTag=warm' \
  -F 'sampleAudio=@./sample.wav'
201 Created
{
  "id": "uuid",
  "displayName": "Acme Brand Voice",
  "personalityTag": "warm",
  "providerVoiceId": "vp_abc123",
  "status": "ready",
  "createdAt": "2026-05-13T15:30:00Z"
}

Personality tags

The same voice can be tuned across personalities. We pass the tag to the synthesis prompt, which adjusts pacing and intonation. Available tags: warm, professional, energetic, calm, custom (free-text style guide).

Where the cloned voice plays

  • Smart Hold. The voice that engages callers in the queue.
  • AI receptionist. The voice that handles press-1 calls.
  • Voicemail greetings. Generated TTS messages, per DID.
  • Outbound confirmations. Appointment reminders, callback scheduling.

We require attestation that the person whose voice is cloned has consented in writing. This is in the terms of service for the voice-profiles endpoint. Voices uploaded to clone celebrity, public-figure, or third-party voices will be flagged and removed. See our acceptable use policy.

Quality benchmarks

In our blind A/B testing (40 listeners, 5 brands, 100 utterances each), our cloned voices scored 4.3 out of 5 on perceived authenticity versus original recordings (4.7), and significantly above generic stock voices (3.1). Real-world quality depends heavily on the recording; budget five minutes for a clean take.

Common pitfalls

  • Reverberant room. The model encodes the room into the voice; the clone then sounds boxy. Record in a closet with clothes if you have no booth.
  • Too short a sample. Ten seconds works but adds artifacts. Fifteen to thirty seconds is the sweet spot.
  • Wrong personality tag. A calm voice forced into "energetic" sounds uncanny. Pick the tag that matches the natural delivery of the sample.