Ultra-Low Latency In-Memory Audio Pipeline

Ultra-Low Latency In-Memory Audio Pipeline (Milestone 1: Ingestion & VAD - Milestone 2: Multi-Target TTS) Overview: We are developing a strictly offline, privacy-first audio benchmarking pipeline under Windows. We require an experienced low-level Python / DSP audio engineer to implement Milestone 1 of our system, designed from the ground up to support multi-channel output in Milestone 2. Project Structure & Scope: - Milestone 1 (Current Starter - Fixed Price: $250): Low-latency USB hardware audio capture, native-to-16kHz in-memory resampling, offline Silero VAD chunking, and a thread-safe multi-consumer queue architecture. - Milestone 2 (Follow-up - Fixed Price: $1,250): Single-source transcription (faster-whisper), parallel multi-target text translation (CTranslate2), multi-stream speech synthesis (Piper TTS), domain-specific prompt injection, independent USB soundcard routing (1 output device per target language), and standalone packaging. Milestone 1 Deliverables & Technical Specifications: 1. Hardware Audio Ingestion: Capture 24-bit/48kHz or native streams from an external USB audio interface (e.g., DJI Mic 2) using sounddevice / WASAPI callbacks. 2. In-Memory Resampling: Resample incoming audio strictly in RAM to 16 kHz Mono (NumPy / io.BytesIO) with zero temporary disk writes. 3. Offline Voice Activity Detection (VAD): Segment continuous audio streams using a local Silero VAD model (ONNX Runtime). Silence threshold must be configurable between 450 ms and 600 ms to trigger chunk finalization. 4. Scalable Queue Architecture: Push finalized audio chunks into an in-memory, thread-safe queue structured to seamlessly feed parallel downstream engines in Milestone 2. 5. Robustness & Diagnostics: Automatic stream reconnection on USB disconnect/reconnect, zero buffer overruns, and a lightweight latency logger (RAM-only / console metrics). Strict Technical & Legal Constraints: - 100% Air-Gapped / Offline: Code must run with zero network activity (verified under Windows Airplane Mode). - Zero Disk I/O: Audio data must reside purely in volatile RAM. No temporary .wav or scratch files written to disk. - IP & Licensing: Full Work-for-Hire copyright assignment. Exclusively permissive open-source dependencies (MIT / Apache 2.0). Strict No-GPL policy. Mandatory Application Questions (Please address directly in your proposal): 1. Which library (sounddevice, PyAudio, etc.) do you prefer for low-latency WASAPI USB audio streaming on Windows, and why? 2. Have you previously routed multiple independent audio playback streams to separate physical USB soundcards simultaneously in Python? 3. Briefly outline your approach to ensuring true Zero Disk I/O while handling continuous resampling and VAD chunking in RAM.