Real-Time Audio Capture and Playback Pipeline for Zephyr with Optional ML Integration

Zephyr provides low-level I2S and DMIC driver APIs along with a few narrow samples (echo, codec), but lacks a reusable application-layer audio pipeline that handles real-world complexity. The closest prior effort (GSoC 2025, PR #96657) attempted to integrate TensorFlow Lite Micro Micro Speech with a Linux host streaming setup, but the audio pipeline was tightly coupled with ML logic, limiting extensibility and resulting in fragile real-time behavior. A clean abstraction to support multiple backends is still absent. This proposal addresses these gaps by introducing a modular, backend-agnostic audio pipeline, designed around fixed-size PCM blocks and explicit ownership semantics. The system will support continuous audio streaming at 16 - 48 kHz, 16-bit PCM, targeting 4 - 10 ms frame latency with zero TX underruns and zero RX overruns under sustained load. Key deliverables include a pipeline API, support for multiple backends (I2S, PDM, optional USB), and reusable processing stages such as format conversion, and ML inference as an upstream sample.