On-Device Audio Decision Pipeline (Ear + Voz + Jev)

MEDIA PREVIEW

Overview & Result

Desert Ant Labs revealed a high-performance local audio pipeline designed for embedded and privacy-critical environments. Combining on-device voice activity detection (Ear), local streaming transcription (Voz), and PII scrubbing (Redact), the pipeline uses Jev as its central high-speed decision engine, executing 20 intent and routing choices in under 1.3 seconds total without transmitting raw audio to third-party cloud LLMs.

How Jev fits in the loop

1. Ear library performs sub-10ms voice activity detection on local microphone input.
2. Voz generates on-device transcription with zero external network overhead.
3. Redact masks sensitive personal data (names, credit cards, addresses) in memory.
4. Cleaned text is piped into Jev API for immediate intent determination, device command execution, and permission verification.

How to reproduce

1. Clone desertantlabs/edge-audio-pipeline and install Rust toolchain (cargo build --release).
2. Run the local audio daemon listening on system audio input.
3. Speak compound commands (e.g., 'turn off living room lights but do not change the thermostat').
4. Inspect stdout logs confirming sub-70ms sequential Jev intent choices and PII redaction masks.

Why this build matters

Proves that modern voice agents don't need cloud-based multi-billion parameter LLMs for decision making. By pairing local speech models with Jev, agents achieve unprecedented speed, privacy, and low power consumption.

Reported performance

Cost: $0.0002 / interaction

Latency: 65ms / decision

Limitations

• Requires local CPU/NPU capable of running quantized Whisper/Voz speech models in real-time.
• Network connection required for Jev cloud decision API (or local Jev edge container when available).

Patterns