Using Cepstral Text-to-Speech for Per-Lead Voicemail

A single canned voicemail works, but a message that says the customer's name and references their account lands better. Cepstral brings text-to-speech (TTS (text to speech)) to VICIdial, which lets you build per-lead answering-machine messages from data instead of recording one static prompt for everyone. When answering machine detection (AMD (answering machine detection)) decides a machine picked up, the message it leaves can be generated on the fly.

What Cepstral adds

With the integration in place, any of your campaign-related audio prompts can use TTS scripts rather than fixed audio files. Cepstral uses SSML, so you control pronunciation, volume, pitch, and rate inside the script. The speech scripting can also pull from the default lead tables, which is what makes per-lead messages possible: the same script produces a different spoken message for each Lead.

In practice that means the Answering Machine Message you would normally point at a recorded file can instead point at a TTS script that weaves in fields from the Lead list. One script, thousands of tailored voicemails. A message that opens with the right name and a real account reference sounds like a callback worth returning, not a blast, and that lift in callbacks is the whole reason to go to the trouble.

How a per-lead message is built

flowchart TD
  A[Machine detected by AMD] --> B[Campaign uses TTS script]
  B --> C[Script reads lead fields]
  C --> D[Cepstral generates speech with SSML]
  D --> E[Save to file]
  E --> F[Play personalized message to machine]

The script reads the lead's data, Cepstral renders it to speech with your SSML settings, and the audio is saved so VICIdial can play it. That save-to-file step is not optional plumbing; it is a core part of how the TTS process feeds the voicemail.

The three licenses you need

Before buying, work out how many concurrent channels you need and how many dialers will run it, because the TTS software installs on each dialer that uses it. The integration needs three separate licenses:

1. Voice license, per server. Pick a speaking voice that fits the campaign, and make sure it is the 8kHz Linux build.
2. Channel license, per channel. This caps how many lines can speak at the same time.
3. Save-to-file license, per server. This is required for VICIdial's TTS process to work at all.

Size the channel count to your busiest moment. If more lines need to speak concurrently than you have channels for, messages queue, and a queued voicemail is a missed one. Count the dialers too, since the software installs on each one that will speak, and the voice you pick should match the campaign's tone. A formal collections message and a friendly appointment reminder rarely want the same voice.

There is a real trade-off against plain recordings. A recorded prompt is one fixed file with no per-call cost and nothing to render; TTS spends a moment generating audio for every machine it leaves a message on. For a generic message the recording wins on simplicity. The moment you want the customer's own details spoken back to them, TTS is the only practical way to do it at scale, and the save-to-file step means each rendered message becomes a real audio file VICIdial plays like any other.

Where TTS fits in your AMD flow

TTS only matters once you have decided detected machines should hear a message rather than a hangup. For how that decision works, read about when to use VICIdial AMD, and for the full detection picture see the AMD and CPD complete guide. Keep your scripts short and verify pronunciation of names before a big run.

We help size channels, install Cepstral across your dialers, and build the lead-aware scripts. See VICIfast pricing to add personalized voicemail to your campaigns.