{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/233b6074ec874490b124afb2227c6dda\" frameborder=\"0\" width=\"1920\" height=\"1440\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1440,"width":1920,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1440,"thumbnail_width":1920,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/233b6074ec874490b124afb2227c6dda-f7cfc7abf2ddf2c9.gif","duration":670.054,"title":"Kyvera: Stress-Testing Voice Agents in Real Conditions","description":"This Loom demonstrates Cabra, a platform for benchmarking and stress testing voice agents under realistic conditions before production deployment. It generates natural call dialogues, adds filler words, converts the text to audio, applies a telephonic codec, adds background noise, and streams the degraded audio to the agent to test speech-to-text and escalation behavior. In a demo scenario with a frustrated caller complaining about a deducted amount, the agent scored 23% overall and produced garbled ASR output due to language mixing and noise. The trace identified key failures: it did not collect a required account number despite the user prompt, and it escalated to a human even when escalation was instructed not to occur."}