In cascaded voice agents, automatic speech recognition (ASR) errors do not all matter equally. A small transcription mistake may have no effect on the tool call that follows, while another may change the call's intent or its arguments. The distinction is important for building reliable voice interfaces, but standard metrics are poorly suited to capturing it.
Conventional word error rate (WER) counts every transcription error the same way, regardless of what the agent does next. The authors of this paper argue that this makes WER a misleading proxy for real-world performance in tool-using voice agents. They propose task-role diagnostics that classify ASR errors by their downstream consequences, offering a more meaningful way to evaluate and improve these systems.