Full-Duplex Voice Research Lab
Standard conversational pipelines wait for awkward seconds of silence before an LLM even begins generating. YuktYantra researches native speech-to-speech models that listen, speak, and backchannel concurrently every 80ms on GPU infrastructure.
Serving Target: sub-80ms Frame Baseline (Design & Profiling Phase)Live Full-Duplex Frame Stream (80 ms Cadence)
12.5 FPS · Real-Time PlayheadContinuous Bidirectional Interaction
Why turn-based pipelines break conversational realism, and how duplex architectures target human latency budgets.
Legacy Cascaded Pipeline
Latency compounds across each independent stage. The mandatory silence detection timeout is often longer than natural turn-taking thresholds (200-300ms), making the bot feel unresponsive and robotic.
Native Full-Duplex Engine
The model receives raw audio tokens continuously. It autonomously predicts whether to speak, listen, backchannel (e.g., "haan", "okay"), or gracefully pause mid-syllable when interrupted.
Our Optimization Architecture
Batched Graph Execution
Cross-session tensor batching under unified 80ms ticks. CUDA graphs locked to static shapes eliminate kernel launch gaps. Attention sink sliding-windows prevent unbounded memory growth.
8kHz Channel Adaptation
Most foundation models train on 24kHz wideband audio. We evaluate acoustic degradation over Indian narrowband telecom networks (AMR / G.711) and deploy lightweight pre-model normalizers.
Language & Accent Adapters
Building atop open Hindi baselines (Human-1) to support Kannada, Tamil, Telugu, and conversational Hinglish code-mixing using synthetic dual-channel dialogue datasets.