Notes from the Edge

How we built a 38M-parameter speech model that runs in real time on the silicon our customers already have.

ABR’s Nith family brings natural, production-ready text-to-speech to edge devices at just 5.5 million parameters. The article explores how Nith delivers low-latency, real-time voice generation on-device while supporting features like voice cloning, prosody control, multilingual output, and a compact memory footprint.

Niagara-38m sets a new benchmark for edge speech recognition, delivering 90.64% accuracy at just 38 million parameters. The article explores how ABR’s State Space Model architecture enables real-time, on-device ASR with low latency, high efficiency, and performance competitive with models many times larger.

Cloud voice AI can cost far more than teams expect. This article breaks down how ASR, TTS, audio tokenization, and voice-first model behaviour compound cloud inference costs, and why moving speech processing on-device can dramatically improve the economics of voice AI.

Artificial intelligence is rapidly moving from the cloud to the device. Phones, wearables, industrial sensors, and embedded systems increasingly need to run AI models locally for reasons of latency, privacy, reliability, and cost.