A bilingual voice agent that answers in English or Urdu, on CPU alone.
A real-time reception kiosk for a datacenter inauguration, running the full speech-to-speech stack on CPU-only compute. Sentence-level streaming cut first-audio delay from over 7 seconds to a sub-second budget.
Industry
Datacenter / Telecom
Platform
Voice agent, kiosk
Scope
Real-time bilingual voice AI
Services
Voice AI, Latency Engineering
Overview
The client needed a reception kiosk that answers questions in English or Urdu from one internal reference document, running entirely on CPU with no GPU available. Two text-to-speech pipelines run inside a single conversation loop, one for each language.
The challenge
The first version waited for the full LLM response before speaking, so the first audio took between 7 and 12 seconds. Scarce CPU cycles, not model quality, were the bottleneck, which ruled out simply adding parallel compute.
Our approach
We switched to sentence-level streaming so speech starts on the first sentence, dropped the TTS frame size from 200ms to 20ms, and used Silero VAD for accurate turn-taking. Stage-level tracing in Langfuse showed where the time was actually going.
What we designed
A speech-to-speech loop tuned for a CPU-only deployment.
Bilingual speech loop
LiveKit Agents, Whisper transcription, and Qwen over vLLM, with Kokoro for English and MMS-TTS for Urdu synthesis in one conversation.
Sentence-level streaming
Synthesis starts on the first sentence instead of the full response, cutting first-audio delay from 7 to 12 seconds down to a sub-second budget.
Stage-level observability
Langfuse spans split VAD, STT, retrieval, generation, and TTS, surfacing tts_first_frame as the metric that mattered.
Inside the kiosk
The conversation people see, and the traces that made it fast.
Kiosk experience
2 screens
A bilingual welcome that answers from a single reference document.
Welcome screen
Live bilingual conversation
Pipeline and tracing
4 screens
The CPU-only stack and the observability that tuned it.
Language routing
Grounded answer with source
Langfuse latency trace
Pipeline observability
Selected Work
Case studies we're proud of.
A look at products we designed, built, and shipped with teams who trusted us to get it right.
CLIENT VOICES
DEV N SCALE felt less like an agency and more like the most senior people on our team — they shipped exactly what we needed, on time, and pushed back when it mattered. — Lukas Meyer, VP Product, Ledgerly
FAQ
Questions,
answered.
Can't find what you're looking for? Reach out and we'll get back to you quickly.
We provide end-to-end software development services, including custom web applications, mobile app development, UI/UX design, cloud solutions, AI-powered applications, system integrations, and ongoing maintenance and support.
