Retail Multilingual AI Voice Translation Customer Engagement

Multilingual Customer Engagement Platform

A real-time Japanese–English AI translation service supporting voice and text, designed for retail marketing and customer acquisition, with strict session controls, cost governance, and production-grade security.

2-way

Real-time voice and text translation (Japanese ↔ English)

<300ms

Token-to-connection latency via direct browser WebSocket

Live

Production-deployed, publicly accessible at spaxialiq.ai

🗾 Japanese ↔ English 🎙 Voice + Text Translation 🔒 Server-Side API Security ⏱ Session-Controlled

The Challenge

Language Barriers Limiting Customer Reach

A retail company with a significant Japanese customer segment needed a way to bridge communication gaps in real time. Existing translation tools were either slow, text-only, or required customers to navigate third-party apps, creating friction at exactly the moment when engagement matters most.

The requirements were specific: voice and text translation, bidirectional (Japanese to English and English to Japanese), low enough latency to feel natural, and secure enough to deploy as a public-facing demo without API key exposure risks.

Cost control was equally important. A public-facing translation service without usage limits could generate unpredictable API costs. The solution needed hard session limits, not just soft guidelines.

The Solution

Real-Time Translation with Session Control and Secure Architecture

I extended the existing voice AI platform with a dedicated translation service endpoint. The architecture preserves the low-latency properties of the core platform while adding the translation-specific controls needed for a retail use case.

Browser (user speaks in Japanese or English)
    ↓ POST /translation-service-session
Backend (Express / Node.js)
    → Verify Firebase anonymous auth
    → Check daily session quota (2 sessions × 5 min max)
    → Select system prompt (Strict or Conversational mode)
    → Request ephemeral token from xAI (TTL = 300s)
    ← Return token + session metadata (API key never leaves server)
Browser connects directly to xAI WebSocket using ephemeral token
    ↓ Real-time bidirectional audio stream
AI translates speech → returns translated voice + text output
Client enforces session timer: 240s warning → 300s disable → 305s close

The key architectural decision: the browser connects directly to the AI voice API using a short-lived ephemeral token, rather than routing audio through the backend. This preserves sub-300ms latency while keeping the actual API key entirely server-side.

Strict Translation Mode

Deterministic translation only
No explanations or commentary
Rejects off-topic input
Enforces translation-only boundaries

Conversational Translation Mode

Allows tone adjustments and iteration
Supports refinement requests
Maintains translation-only scope
More natural for extended conversations

Security & Cost Control by Design

System prompts are constructed server-side and bound during ephemeral token creation and cannot be tampered with by the client. The daily quota (2 sessions × 5 minutes) is enforced server-side and on the client simultaneously. Hard session closure at 305 seconds prevents any possibility of runaway API usage. The API key is never visible to the browser at any point in the flow.

React Express / Node.js TypeScript xAI Realtime API Firebase Auth WebSocket GCP Cloud Run Firebase Hosting

The Results

Low-Latency Translation, Production-Ready from Launch

The translation platform is live at spaxialiq.ai/translation and has been used successfully for retail marketing demonstrations.

Real-time

Voice and text translation with natural conversational latency