F2 · Edge AI & Agents
DGX Spark Voice Chatbot (VAD/ASR)
Overview
On an NVIDIA DGX Spark, this guide chains a USB microphone, PyAudio, WebRTC VAD, turn detection, faster-whisper and vLLM into a fully local voice-to-text-response pipeline. The Grace CPU handles audio capture, segmentation and speech-to-text while the Blackwell GPU runs LLM generation, validating segmented transcripts and text responses from live voice input. This hybrid workflow chains open-source audio processing and local LLM serving into an offline voice-application baseline where developers can tune model, segmentation and prompt strategy.
Scope of verification
- It is an offline voice-chatbot reference workflow on a single DGX Spark, validating end-to-end integration of microphone input, VAD/turn detection, CPU speech-to-text and GPU-backed text generation.
- Dependencies include the Grace Blackwell GB10, a USB microphone, PyAudio/WebRTC VAD, faster-whisper, vLLM, the specified models and a stable local audio environment; results vary with accent, noise and segmentation parameters.
- Scope centers on voice input, transcription and text response, with no TTS; it also does not cover consistent latency/ASR-quality benchmarks, barge-in, far-field audio, long-run reliability or production voice-service operations.
Canonical source: https://learn.arm.com/learning-paths/laptops-and-desktops/dgx_spark_voicechatbot/
This page is a guided summary with boundary notes, not an official reproduction.
This page is a guided summary with boundary notes, not an official reproduction.