duplexjev

DuplexJev is a specialized System One speech decision model designed to act as the ultra-fast perception layer for full-duplex voice agents. Developed by Adventists AI, it bypasses traditional speech-to-text transcription to answer multiple categorical questions about raw audio in a single forward pass.

For the Non-Technical Reader

When humans converse, we do not wait to transcribe someone's words before deciding if they are interrupted, angry, or taking a breath—our brains react instinctively. DuplexJev gives AI voice assistants these exact same reflexes. Instead of waiting for a slow transcription and processing loop, it listens to raw audio tone, pauses, and pitch to instantly know when to speak, when to yield, or if the user sounds frustrated. This eliminates awkward dead air and makes conversational AI feel natural and fluid.

For the Technical Reader

DuplexJev replaces the traditional ASR-to-LLM pipeline with a single-token readout mechanism built on a frozen speech LLM served via vLLM. Key technical specifications include:

• Speed & Latency: Evaluates 10 distinct decisions in just 92 ms on a single NVIDIA H200 GPU—21x faster than running an equivalent ASR-to-LLM stack.
• Architecture & Dynamic Prompting: Operates as a System One model reading single-token logits per question without autoregressive text decoding. New decision categories can be added at runtime via single-line text prompts without model retraining.
• State-of-the-Art Benchmarks: Outperforms Gemini 3.8 Flash, Gemini 3.8 Live, GPT-realtime-2.1, and Qwen3-Omni across public paralinguistic benchmarks (LISTEN, StressTest, and MSPB).
• Multimodal Capabilities: The DuplexJev-32B-Vision variant cross-references visual inputs with audio signals in the same pass (e.g., matching a spoken stressed word against image regions).
• Hardware & Cost: DuplexJev-4B requires ~10 GB VRAM with operational costs under $0.50 per million decisions, while DuplexJev-32B fits on a single 80 GB GPU.

Why It Matters

By decoupling low-latency conversational control (System One) from high-level LLM reasoning (System Two), DuplexJev fundamentally changes the economics and architecture of voice applications. It dramatically lowers compute costs while solving the latency and barge-in bottleneck that plagues current voice agents, making ultra-responsive open-source deployments accessible.

The Voice AI Space Lab Idea

The Real-Time High-Stakes Negotiation Simulator: Build an interactive training platform for executive negotiators or customer service teams. Powered by DuplexJev, the system instantly analyzes pitch, micro-hesitations, tone shifts, and background pauses to trigger real-time emotional reactions or mid-sentence counter-arguments the exact millisecond a trainee hesitates.

Explore the project on the DuplexJev GitHub Repository or test the interactive experience via the DuplexJev Live Demo.