Open Research
Our research output — papers, technical reports, and datasets. Everything is open-access. Nothing requires an institutional login.
Research in progress
CogniHuman is an early-stage research foundation. Our first publications are currently being prepared. Below you can see what is in the pipeline. We will link papers and datasets here as they are released.
A systematic latency breakdown of end-to-end voice agent pipelines (ASR → LLM → TTS → VAD) running on CPU-only hardware. We introduce a standardized profiling methodology and identify primary bottlenecks across pipeline stages. Includes cross-platform measurements on Windows, Linux, and macOS.
We introduce VoiceBench, a reproducible benchmark suite for evaluating on-device voice agents across latency, accuracy, and resource consumption. VoiceBench provides standardized test sets, evaluation protocols, and baseline measurements across diverse hardware configurations.
A systematic study of quantization trade-offs (Q2, Q3, Q4, Q8) for voice pipeline components. We evaluate the effect of different quantization strategies on perplexity, WER, naturalness (MOS), and inference latency across CPU and edge hardware.
All CogniHuman papers will be published open-access on arXiv and linked here. Datasets will be released on Hugging Face under CC BY-SA 4.0. Code on GitHub under Apache 2.0.