Blogs›Decagon›Voice Activity Detection (VAD) and Turn Detection
Decagon has developed novel statistical methods to improve turn detection in voice AI pipelines. This includes addressing miscalibrated probabilities in Voice Activity Detection (VAD) models using isotonic regression and replacing heuristic-based turn ending decisions with a Bayesian hazard statistic. These improvements have led to significant reductions in broken turns and transcription errors, enhancing the reliability of voice AI agents in production environments. The company is also focusing on scaling voice agent development by transforming tribal knowledge into playbooks, automating manual configurations into product features, and leveraging simulations to reduce manual testing. This has resulted in a ~37% reduction in time-to-go-live and a halving of engineering hours per launch, enabling customers to build sophisticated outbound flows independently.
7/13/2026
This post details the efforts to scale voice agent development by codifying tribal knowledge into playbooks, converting manual configurations into product features, and implementing Duet simulations for testing. These initiatives have led to a ~37% reduction in time-to-go-live and a significant decrease in engineering hours per launch, empowering customers to build complex outbound flows autonomously.
4/9/2026
This post details principles for improving voice agent conversational design and voice profile selection. It outlines five principles for good conversational design: helping users get things done fast, sounding clear and natural, recovering when things go wrong, showing empathy, and knowing your users. It also discusses choosing a voice profile by aligning TTS voice dimensions (formal/casual, energetic/calm, mature/youthful, warm/professional) with persona guidelines and emphasizes the importance of A/B testing and data validation for both conversational design and voice profiles.
3/24/2026
This post introduces a new approach to improving turn detection in voice AI pipelines by addressing two key issues: miscalibrated VAD probabilities and a flawed decision framework for turn endings. The authors applied isotonic regression to recalibrate VAD model outputs, significantly reducing calibration error. They also developed a Bayesian hazard statistic that replaces heuristic-based silence detection and timeouts with a learned statistic that accumulates evidence of a true turn end. This new statistic smooths out within-speech dips and responds decisively to genuine silence, leading to a 42% reduction in broken turns and a 20% reduction in pipeline Word Error Rate (WER).
3/9/2026
This post introduces Decagon's outbound voice capabilities, detailing the technical challenges and solutions for proactive customer engagement. It highlights the use of Agent Operating Procedures (AOPs) for designing scalable outbound workflows, intelligent and compliant calling mechanisms (including voicemail detection, robocall screener navigation, low-latency turn detection, and retry strategies), and the capture of customer intent for personalized engagement. The system leverages fine-tuned models for faster AOP selection and execution, and features like branded calling identities to increase trust and answer rates. Missions are introduced for managing outbound campaigns with batch dialing and customized follow-ups.