Research
We strive to find where voice AI breaks today using real human data.
Library
We regularly publish and open-source audio data for research purpose. We focus on where voice and audio models break today. If you are interested in collaborating, please email partnership@besimple.ai.
September 10, 2026
A full-duplex voice benchmark for measuring whether an ongoing speaker continues, adapts, or yields when a listener contributes during the turn.
Can a voice agent incorporate a listener's correction or completion while it keeps speaking, rather than ignoring the cue or yielding the floor?
August 28, 2026
A controlled scaling study of targeted human speech data for improving exact structured-value transcription in Inkling.
How much can targeted human speech data move an already capable speech model on the exact values that voice workflows depend on?
July 7, 2026
A vocal emotion benchmark for evaluating whether emotion detection or omni models can identify expressed affect from raw speech without transcripts or metadata.
Voice agents often rely on a separate model for emotion detection, but how accurate are they in detection human emotions?
May 1, 2026
A speech-to-text benchmark for exact structured values in English workplace speech, including emails, command-line flags, file paths, URLs, account identifiers, dates, and measurements.
Voice agents can transcribe something incorrectly but still get to the right outcome, but what if it's on critical tokens that cannot be recovered later?
