Built to be heard.
A prospect's name, title and company go in. About ten seconds later, an English or French voicemail comes out, written by an LLM and spoken in a cloned voice, ready to play mid-call.
The brief
A four-stage pipeline (draft, synthesize, deliver, scale) that turns a name, title and company into a bilingual voicemail in a cloned voice. It plays in the browser in about ten seconds, or goes out by email after an explicit confirmation.
Federally regulated and Quebec-facing accounts expect service in both official languages, which usually means a bilingual hire or a studio session. This turns it into one extra API call, so the same rep can cover those accounts the same day without adding headcount.
Hear it
The same prospect run through the same pipeline in both languages. No password needed for these two.
Marie Lavoie, Director of Digital Services at a Canadian bank. Both samples were written and spoken by the pipeline below, one call per language.
Why bilingual, specifically
Federally regulated Canadian institutions, and increasingly any business operating in Quebec under Bill 96, have to serve customers in both official languages. Today that usually costs a bilingual hire, a translation pass, or a second studio session per script. With this tool, an English-speaking rep can leave a fluent French voicemail for a Quebec account the same day.
What changes for a rep
It's built for accounts where fast follow-up and language coverage both matter: federally regulated institutions, Quebec-facing businesses, and anywhere a same-day bilingual voicemail beats a callback that never happens. For a rep, the change is fewer follow-ups lost between calls. A name goes into a form while the last call is still fresh, and a personalized voicemail reaches the prospect before the next call starts.
Enter the password to open the tool
This opens the working generator. Type in a company, prospect and title, and it writes, voices and plays the voicemail back in your browser.
An interactive preview of the CSV, Slack and WhatsApp items below. WhatsApp is waiting on Meta template approval; the others aren't in the web tool yet.
This started as a rehearsed demo: two languages, two hand-approved scripts, nothing generated live. It changed when the question went from "can we generate a good voicemail?" to "can we generate one for whoever's on the other end of this call, right now?"
You type a name into a form while you're still talking. Before you've finished the sentence, the room hears a voicemail in your voice for a company you first heard of ninety seconds ago.
That's the demo this tool was built for. The rest of this page explains how it works.
One pipeline, one branch point
What each stage does
Draft
An LLM call takes company, prospect and title and writes a ~70-word script sized for voicemail, in the right language, ending with the callback number spelled out digit by digit so it's easy to catch by ear.
Synthesize
The script is rendered through a Professional Voice Clone, on the model that clone was fine-tuned for. That's the correct model, which isn't always the newest one.
Deliver
The mp3 either plays immediately in the browser or goes out as an email attachment. The email path always shows the recipient, the content and the file size before it sends anything. Cold commercial email in Canada falls under CASL, the same law that shapes the live-dialing plans below. This demo's gate confirms you want to send; it doesn't yet cover full CASL compliance like sender ID and unsubscribe.
Scale
The same three stages run over a CSV of prospects instead of a single form. Every script is shown before any credits are spent on audio, and every recipient is shown before any email goes out. Nothing is sent without an explicit yes.
Two ways to run it
The tool ships as two separate paths because they carry different risks.
The demo pipeline
Rehearsed
Two scripts, one per language, written and approved before anyone heard them. Nothing is generated at demo time, so a bad line gets caught in rehearsal instead of in front of an audience.
The live pipeline
On-demand
Written by an LLM for each prospect in seconds, and not reviewed before it's spoken. To make up for that, nothing is sent to another person without an explicit confirmation.
Dead ends, caught before they cost anything
Each of these delivery channels sounded plausible and didn't hold up once I read the docs and terms of service.
Doesn't exist. The Twilio integration only handles live phone calls. It can't deliver a file over any channel.
Against the rules. LinkedIn's API has no third-party DM endpoint; automating it via the UI violates the User Agreement.
Wrong platform, wrong shape. OpenAI's iMessage plugin is a macOS desktop feature with no API a server can call.
Real, but gated. Twilio can do it, but only with a Meta-approved message template, which is an approvals process rather than code.
Same pipeline, more ways out
The draft → synthesize core doesn't change. What changes is who triggers it and where the result goes.
CSV bulk upload
Already works from the command line, one row or a whole prospect list at a time, with two review gates so nothing is sent unseen.
Slack as a second front door
Too slow to replace the live-call form, but useful for teammates who want to queue up voicemails ahead of a call.
WhatsApp delivery
Buildable on Twilio's API today, but blocked until Meta approves an outreach template.
Live outbound dialing
Having the clone place the call itself means switching to ElevenLabs' Conversational AI product, and running into robocalling law whenever the person hasn't opted in first.
iMessage delivery
OpenAI's ChatGPT plugin can send iMessages, but it's a macOS desktop feature with no API a server can call.
Preview the roadmap → to click through the CSV, Slack and WhatsApp items above.
Built to be heard
One rep, one cloned voice, and a few cents per prospect instead of a studio session.