Voxbench

Voice-agent teams change speech, language, turn-taking and telephony components frequently, yet a conventional unit test cannot reveal interruptions, delayed responses, voicemail errors or conversational completion. Voxbench proposes a hosted synthetic-call harness that replays frozen scenarios and compares P50 and P95 latency, false interruption, turn detection, voicemail detection, completion and disclosure-presence signals across builds. The supplied research confirms several active voice-agent evaluation products, including two with published self-service entry points. That validates the category while weakening the original claim that available tools are uniformly enterprise-only.

Synthetic audio is controlled evidence, not representative production traffic. Scenario wording, audio codec, noise, carrier path, geography, language, voice and measurement clock all affect results. P50 and P95 are meaningless without sample size and a defined start and stop event. A detected disclosure phrase does not prove legally sufficient timing, wording or consent. A completion label is not a successful customer outcome. A failing test should propose a gate result under an owner-approved policy; it should not silently block a deployment or merge without explicit repository and release authority.

Scenario source, corpus version, expected behavior, test authorization, call attempt, carrier event, received audio, measured event, metric calculation, threshold version, regression candidate, reviewer finding, gate decision, deployment action, production observation and business outcome remain separate. The first release should prove reproducibility across one verified endpoint and a narrow metric set before claiming vendor neutrality, standard benchmarks or a durable data moat.

An engineering or quality lead shipping a voice agent frequently enough that provider changes and conversational regressions create repeated manual testing work.