Voice commands in four languages

The chat panel has a microphone icon for a reason: typing geographic queries on a phone in a Land Cruiser is a bad time. Press-to-talk works in English, Turkish, Russian, and Uzbek, with on-device Web Speech transcription so your voice never leaves the browser.

How to use it

1. Open the chat panel (mascot icon, bottom-right). Pick your language from the language selector if it isn't right.
2. Press & hold the mic icon, speak your question, release.
3. Wait for the transcript. The text appears in the input box — review it, edit if needed, send.

What gets routed where

Behind the scenes your transcript hits the P1 Main Orchestrator (Haiku 4.5), which classifies intent and decides which agent should handle it:

• "Show me parcels in Andijan" → Haiku 4.5 fields it directly with a single MCP tool call.
• "Translate this report to Russian" → P3 Report Translator (Sonnet 4.6, section-by-section).
• "Is this 412-ha parcel suitable for Pistacia vera?" → P2 Area Pipeline (Opus 4.7 + 4 specialists).

You can hear the cascade in the latency: Haiku replies in under a second, Sonnet within five, Opus streams over 15-30 seconds.

Language nuances

• Turkish. Best transcription quality after English. Forestry vocab works well: fidan, orman, arazi, sulama.
• Russian. Good for general queries; we noticed Sentinel/MODIS terms get romanised.
• Uzbek. Latin script and Cyrillic both supported. Place-name accuracy is strongest for viloyat-level (Andijon, Farg'ona, Toshkent); tuman-level depends on dictionary coverage.

Privacy

Speech-to-text runs in the browser via the Web Speech API — audio never reaches our servers. Only the transcript is sent to the backend, with the same auth and rate-limit protections as a typed query.