sonar.
Get API key

sonar

[ voice models for products ]

Low-latency speech in, speech out. Runs in your region, on your data, inside your product.

Start buildingRead the docs

move the cursor — the bundle listens

Enterprise-grade infrastructure for voice

01 · Streaming

Speech starts answering before you finish the sentence.

Audio is chunked at 40 ms and decoded as it lands. The first token leaves our edge in under 180 ms; the rest streams as one field.

180 msto first token

02 · Regions

Runs where your users are, never anywhere else.

Pick Frankfurt, Zagreb or Sarajevo at request time. Audio never leaves the region, and nothing is retained after the socket closes.

3EU regions

03 · Grounding

Every answer is tied to a document you gave it.

Index your manuals, menus, timetables. sonar cites the passage it used and refuses gracefully when nothing matches.

100%cited answers

Four models.
One endpoint.

Prices per minute of audio, billed per second. Latency is p50 from Sarajevo edge.

ModelLatencyLanguagesPrice / minBest for
sonar-1 mini120 ms12€0.004Dictation, commands, short replies.
sonar-1[ Default ]180 ms31€0.011Assistants, support lines, in-app help.
sonar-1 pro260 ms31€0.024Long dialogue, tools, grounding.
sonar-echo[ TTS ]90 ms18€0.006Text to speech, six studio voices.

[ Try ]

Ask it something you would ask a colleague.

This is the real response shape: a transcript line, then the answer streamed word by word, then the source it used.

sonar · terminalv1.4

$ sonar ask --model sonar-1 --region sa

source: gras-timetable-2026.pdf · p.14 · confidence 0.94

[ AI Product ]

Less typing.
More talking to the product.

sonar ships as one endpoint, one SDK and one bill. No GPUs to rent, no models to babysit.

Changelog

  1. 1.4

    02 Sep 2026

    Sarajevo region, Bosnian dialect pack

    • New sa region on the Skenderija edge.
    • Ijekavian and ekavian handled by one model.
    • Streaming resumes after a dropped socket.

  2. 1.3

    18 Jul 2026

    Grounding cites page numbers

    • PDF sources return page and line.
    • Refusal message is now configurable per app.

  3. 1.2

    03 Jun 2026

    sonar-echo, six voices

    • Text to speech with 90 ms first audio.
    • Voices tuned for phone lines and earbuds.

  4. 1.1

    21 Apr 2026

    Tools in dialogue

    • Call your functions mid-sentence.
    • Timeouts fall back to a spoken apology.

[ Docs ]

Three lines to first sentence.

1import { Sonar } from "@sonar/sdk";2const s = new Sonar({ region: "sa" });3await s.talk(mic, speaker, { model: "sonar-1" });
Open the docs