TurnBench evaluates turn-taking in spoken dialogue
Open benchmark with 30 hours of dyadic speech annotated by three people, a scoring protocol based on conversation analysis, and results from 14 systems for measuring turn-taking.
Source: TurnBench avalia a alternância de turnos em diálogos falados (sesame.com). Text prepared with AI from this source.
What happened and what to do
The open TurnBench benchmark gathers 30 hours of dyadic speech annotated by three people and proposes a scoring protocol based on conversation analysis. It reports results from 14 systems on measuring turn-taking in spoken dialogue. The focus is on the end of turns and on backchannel signals such as "mm-hmm".
In practice, a company running voice assistants, automated contact-centre lines or conversational transcription products can use the protocol to check whether its systems cut users off at the wrong moment, respond too slowly or mishandle acknowledgement signals. This requires an evaluation set built from the company's own recordings, annotated under similar criteria, and a pipeline that runs the tests on every model or configuration change, with metrics comparable over time.
How the consultancy can help
We can diagnose where your company's voice systems fail at turn-taking, build an evaluation set from your own recordings and criteria, and implement a monitoring pipeline on your infrastructure, with ongoing operation.
Next step
Send a short description of your case, including the voice channel you use and the approximate conversation volume, and receive a scoped proposal from Wendelmaques.
Let's assess your project
I reply with scope and a proposal.