Skip to content

From a few words to a clip you can check.

The steps behind every reply, and the spoken yes that comes before any voice is cloned.

From your voice to a signed answer.

Follow one question through 8 steps. The caller hears the first sound in 0.84 s, and the first sentence of the answer in 2.23 s. The two marked trust steps are the AI disclosure, and the watermark with a signed receipt.

Example, not a real caller: a clinic agent is asked, “Kal Dr. Mehta ki OPD kitne baje hai?” It finds the passage “OPD · Dr. Mehta · Mon-Sat 10:00-13:00” in the clinic's PDF and answers, “Dr. Mehta ki OPD kal subah 10 se 1 baje tak hai.”

Time since the request was sent 3.02 s

Stages overlap: speech starts while the reply is still being written.

  1. At the start of every call or chat the agent tells the caller it is an AI. In our test this spoken line is the first sound back, at 0.84 s, and it plays while the answer is still being written.

  2. Hold to talk in the browser, or take a phone call. Phone audio is 8 kHz, through Exotel (tested on a trial account).

  3. Whisper turns your speech into text, on Groq or on your Mac's GPU.

  4. The best-matching passages from your PDFs are found, even when you ask in Hindi.

  5. A language model writes the reply from those passages only, or says it does not know. By default it is a Qwen3 model on Groq, with gpt-oss models as back-up. With no internet it is a small Llama 3.2 model on your Mac, which is what the times here used.

  6. Kokoro speaks replies for speed, Hindi included. A cloned voice speaks through Qwen3-TTS.

  7. Every spoken sentence gets the inaudible watermark, and a signed receipt is kept. Anyone can check a clip on the Verify page. Check a clip at /verify.

  8. The caller hears the reply in their own language, Hindi and English mixed included. The first sentence of the answer is ready in 2.23 s (median; 3.58 s at p95).

Medians over 30 spoken questions in English and Hindi, with local models on an Apple M4 laptop (16 GB). The clock starts when the audio request is sent, not when the caller stops speaking. Details are on the Proof page.

Voice is personal data. Before Auralis clones anyone, the speaker has to say yes out loud, and Auralis has to hear it.

  1. 1

    Say the sentence

    The speaker reads a fresh sentence and a code aloud. It is new every time, so an old recording cannot be replayed.

  2. 2

    We check it is the same person

    The words and the code are checked, then the voice is matched with the sample. Two different people means no clone.

  3. 3

    Every clip is marked and signed

    An inaudible watermark goes into the audio and a signed receipt is kept for it.

  4. 4

    Anyone can verify it

    Drop a clip on the Verify page. It says whether Auralis made it. No login needed. Open /verify

For developers. Receipts are Ed25519 signatures. The public key is at /.well-known/auralis-receipt-key and the signed withdrawal log, which lists voices whose consent was withdrawn, is at /api/withdrawals.