Talk2Agent
Talk2Agent: Benchmarking Voice Interfaces for Text Agents
A benchmark evaluating voice interfaces by measuring how effectively human-spoken instructions convey task-critical information to LLM-based computer-use agents using real speech recordings.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info talk2agent
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
2 sourced entriesEvery number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Content & Construction: Talk2Agent constructs spoken versions of tasks from WildClawBench (command-line) and OSWorld (graphical desktop). It includes 31.82 hours of human-recorded instructions covering dedicated ASR models, audio-capable LLMs, contextual biasing, and ontology repair.
Scoring & Key Findings: The benchmark proposes an execution-free evaluation framework that projects the original task grader onto prompt-addressable intentions to measure information retention without costly agent re-execution. This native projection correlates with downstream performance (Pearson correlation improvement of 0.246 over WER/CER).