Technology
Go beyond the three‑step pipeline.
Most voice AI runs one transcription model, one LLM and one TTS, or a single speech-to-speech model. ThunderPhone runs many models on every call and lets them check each other.
Architecture
Other stacks break on real calls.
Ours holds up.
Other providers use a three-step pipeline or a speech-to-speech model. Each makes mistakes that ruin real calls.
Each step trusts the last
- A transcript keeps one best guess. When it’s wrong, the LLM can’t recover.
- Fast LLMs make mistakes. Careful ones take seconds to answer.
- Rigid turn-taking makes conversations feel stiff.
Sounds natural, follows rules loosely
- Unreliable on instructions and tool calls.
- Erratic behavior is risky on business calls.
- Sometimes better at hearing, not always.
Many models, checking each other
- Several transcribers plus the audio itself, so callers get heard right.
- A fast and a thinking model, switched turn by turn.
- Tuned by Zeus to what your calls need.
Many models per call
Hear it many ways, then reconcile.
Reliable calls take a lot more than three models. We combine open-source and commercial models, and fine-tune them where it helps.
Several transcribers, plus the audio
We run several transcription models at once, and often send the audio straight to an LLM too.
A fast model and a thinking model
Which one answers is decided turn by turn, based on the conversation.
Noise, speakers, turn-taking
Models for noise, speaker separation and the start and end of each turn.
Natural voice, reliable behavior
We pair speech-to-speech models with text models on the side.
Three-step pipeline wrong answer
ThunderPhone right the first time
Accuracy
Proven on a public benchmark.
Big Bench Audio tests whether a model understands hard spoken questions and answers them correctly.
Our evals
Evals based on real calls.
We test every model for every part of the stack on data from real calls, not vendor claims, and track the pros and cons of each. That’s the signal Zeus uses to pick your setup.
We’ll publish versions of our internal evals soon.
| Real-call eval | Model A | Model B |
|---|---|---|
| Hearing names and numbers | stronger | good |
| Short replies (“yeah”, “no”) | good | stronger |
| Response speed | faster | slower |
| Following long rules | weaker | stronger |
Trade-offs
Every call needs something different. Zeus decides what.
Our architecture is a menu of model options: hearing, thinking, turn-taking and voice, each with options that trade off accuracy, speed and cost. Zeus picks the best setup for your calls.
| Dental receptionist | Outbound sales | |
|---|---|---|
| Hearing | Three transcribers plus audio, for names, birthdays and insurance numbers | One fast transcriber, tuned for short replies |
| Thinking | Thinking model on booking turns, fast model for small talk | Fast model on most turns, to keep the pitch moving |
| Turn-taking | Patient: waits out pauses while callers find a card | Quick: picks up the moment the prospect stops |
| Voice | Stable voice that reads numbers back cleanly | The most natural premium voice |
| Goal | Bookings made, no wrong details | Conversion value minus call cost |
Illustrative examples only. Zeus picks each setup from your own calls and goals.
Latency
Fast on most turns. Careful when it counts.
Some conversation turns need a model to think hard, which takes time. Most can be handled very quickly. ThunderPhone adjusts intelligence turn by turn, so a typical reply comes in about 1.5 seconds, but it takes longer when it needs to, to avoid a mistake. Other voice AI doesn’t.
Response time, turn by turn
Real calls
Natural conversation, on messy calls.
Callers can interrupt, pause and change their minds, and the conversation keeps flowing. ThunderPhone is built for the situations that break other systems.
Mumbly speech
Compares audio and text instead of trusting one transcript.
Names and addresses
Cross-checks proper nouns, spellings, numbers and corrections.
Bad phone audio
Noise models clean up the signal before anything else hears it.
Background chatter
Spots side conversations before they derail the agent.
Mixed-language speech
Switches language mid-call when the caller does.
Interruptions and pauses
Stops when a caller cuts in, ignores “mm-hm”, and waits out a thinking pause.
Voices
TTS voices come with trade-offs. Zeus helps you navigate.
TTS models have different advantages and disadvantages. Zeus can help you pick the best one for your calls.
| Platform Inworld | Premium Cartesia | Premium ElevenLabs v4 Turbo | |
|---|---|---|---|
| Cost | Close to free | About 2¢ a minute more | About 2¢ a minute more |
| Stability | |||
| Naturalness |
Languages
Your callers’ language, mid-call.
Agents understand and speak 62 languages, and switch the moment a caller does. There’s no language surcharge.
Afrikaans
Arabic
Armenian
Assamese
Azerbaijani
Bengali
Bosnian
Bulgarian
Cantonese
Catalan
Chinese
Croatian
Czech
Danish
Dutch
English
Estonian
Filipino
Finnish
French
Galician
German
Greek
Hebrew
Hungarian
Icelandic
Indonesian
Italian
Japanese
Kazakh
Korean
Latvian
Lithuanian
Macedonian
Malay
Nepali
Norwegian
Persian
Polish
Portuguese
Romanian
Russian
Serbian
Slovak
Slovenian
Spanish
Swahili
Swedish
Thai
Turkish
Ukrainian
Urdu
VietnameseTry it on your hardest calls.
Start free with $10 of call credit. No credit card required.