Your agent hears the caller.
MTLM decides what to do.
After speech-to-text, every turn needs a decision: what does the caller want, should a human take over, is this a robocall? MTLM answers with small models on a plain CPU — no LLM call per turn, no GPU, on your own servers.
- Works after any speech-to-text
- French and English
- Runs on CPU, on-prem
- Says when it is unsure
Every turn of a call costs you a decision
An LLM call per turn
Sending each transcript to a large model to pick the next step adds cost on every turn, on every call.
Latency the caller hears
Hundreds of milliseconds to seconds per turn, on top of speech-to-text and speech synthesis. Callers notice the silence.
Handoff at the wrong moment
Angry callers stuck in a loop, or routine calls sent to an operator. Both cost you: churn on one side, staff time on the other.
Try it with a call
Play a sample call, record your own (max 30 s), or paste a transcript. The audio is transcribed by a small Whisper model, then MTLM makes the decision. Everything runs on one shared 6-core CPU, no GPU.
Demo calls are processed then deleted: nothing is stored. Sample voices are synthetic. Live timings vary: the demo box is a shared CPU running other jobs.
Where it plugs in
MTLM sits between your speech-to-text and the rest of your agent. Keep your STT, keep your LLM for the turns that really need one.
Caller
phone, app, web
Your speech-to-text
any model or API
MTLM decision
route · handoff · block
Your agent
LLM only when needed
{"lang": "en", "text": "Can I just talk to a real person please?"}{
"action": "handoff",
"intent": "human",
"intent_confidence": 1.0,
"intent_probabilities": {"booking": 0.0, "change": 0.0, "billing": 0.0,
"complaint": 0.0, "info": 0.0, "human": 1.0},
"p_robocall": 0.37
}What it decides
Caller intent
Booking, change or cancellation, billing, complaint, information — or your own list of intents.
Human handoff
The caller asks for a person, or is angry or in distress: route to an operator before it escalates.
Robocalls and telemarketing
Sales pitches and prize robocalls are blocked before your agent spends a turn on them.
Unsure? It says so
Every answer comes with a confidence. Below your threshold, the agent asks again or a human takes over — no confident guess.
Your own questions
Yes/no questions, a choice between options, or a 1–4 score: urgency, satisfaction, end of call, anything your flow needs.
Your servers
Runs on a plain CPU, in your cloud account or on-prem. Transcripts never leave your network.
Speed, measured
| Tuned to your intents | ~15–20 ms | a small decision model trained on your intents and call logs |
| Untuned engine (this demo) | ~1–2 s | per decision on a quiet 6-core CPU; slower when the box is busy |
| Speech-to-text (this demo) | ~1–4 s | Whisper base for a 10-second call, same CPU — your STT replaces it |
Pricing
Design partner pilot
For the first 3 teams: we tune MTLM on your intents and anonymized transcripts, and measure it against your current setup. In exchange: feedback and a short case study.
Custom decision model
One-off. Trained on your intents, delivered as a small file that runs on your own CPU servers.
On-prem licence
Volume, several languages or call flows, support and updates.
Become a design partner
Frequently asked questions
Which speech-to-text does it work with?
Any. MTLM takes text, so it works after Whisper, your own model or any STT API. The demo uses Whisper tiny/base on CPU only to show the full chain.
Which languages?
French and English are tested in this demo. Other languages are possible; we test them on your calls before promising anything.
How accurate is it?
On our test calls, every automatic decision was correct; unclear calls went to “ask again” or to a human instead. Your calls will differ — that is what the pilot measures, on your own transcripts.
Does it need a GPU?
No. Everything runs on a plain CPU. A tuned decision model answers in ~15–20 ms.
Where does my data go?
Nowhere: in production MTLM runs on your servers. In this demo, audio and transcripts are processed then deleted.
How long to tune it?
A first tuned model takes days, not weeks: your intent list plus a few hundred anonymized transcripts is enough to start.