TechPrufer
Get started
Decagon
Decagon logoDecagon · Feature trail

Voice Activity Detection (VAD) and Turn Detection

6 posts · Feb 2026 – Sep 2026

What this trail covers

Decagon is evolving its voice AI capabilities by focusing on conversational design and voice profiles that reflect brand identity, moving beyond just technical fundamentals like latency and accuracy. The company emphasizes principles of efficient task completion, natural and clear communication, robust error recovery, empathetic interaction, and deep user understanding. This includes aligning Text-to-Speech (TTS) voice characteristics (formal/casual, energetic/calm, mature/youthful, warm/profess). This post details improvements to Voice Activity Detection (VAD) and turn detection by introducing Bayesian methods for improved calibration and a hazard statistic for more robust turn-ending decisions, resulting in fewer broken turns and reduced transcription errors.

All Decagon trails
  1. Audio-native semantic speaker-change detection | DecagonThis post introduces audio-native semantic speaker-change detection as a critical component for voice AI agents. It details the problem of distinguishing relevant speaker changes from background noise or non-interventional speech, and proposes solutions involving speaker embeddings (e.g., TitaNet) and multimodal post-training of LLMs. The post outlines benchmark construction with conversation-level false-positive rate (cFPR) and recall metrics, discusses the limitations of diarization, and explores different methodologies for using speaker embeddings (mean dissimilarity, maxsplit, graph methods). It also details synthetic dataset generation strategies and the supervised fine-tuning (SFT) objective for multimodal models, demonstrating significant improvements in recall at a fixed cFPR.
  2. Why speech-to-speech models aren’t ready for the enterprise (yet) | DecagonThis post details the limitations of current speech-to-speech (S2S) models for enterprise applications, specifically citing issues with reliability (inconsistent behavior, skipped steps), factual accuracy (hallucinations, paraphrasing errors), and cost efficiency (higher compute per audio slice). It contrasts S2S with traditional speech-to-text -> language model -> text-to-speech pipelines, emphasizing the necessity of guardrails (instruction adherence, hallucination detection, escalation) for enterprise workflows. Decagon Voice 2.0 is introduced as a solution that achieves low latency and high reliability through optimizations in generation latency (65% faster), fine-tuned models, streaming generation prosody, and infrastructure improvements (regional deployment), enabling agents to handle complex conversational dynamics without compromising precision and control.
  3. "Agent, agent, agent": designing conversations to reduce barge | DecagonThis post details how Decagon reduced caller barge rates by over 15% in a billing dispute workflow through conversational design changes. The key technical contributions include: 1. Reducing response length to prioritize critical information (e.g., changes in charges since last month). 2. Eliminating system-serving questions and replacing them with pointed questions that advance the caller's resolution goal. 3. Implementing conversational state management to prevent repetition of bill details, thereby shortening conversations and improving caller engagement. These changes were implemented and tested rapidly using Decagon's Agent Operating Procedures and Duet platform.
  4. What I learned building voice agents at Decagon | DecagonThis post details the significant engineering effort and process improvements required to scale voice agent development at Decagon. It highlights the transition from manual, time-consuming development of individual voice agents to a more systematic, productized approach. Key contributions include: transforming tribal knowledge into accessible playbooks for Agent Development Managers (ADMs), productizing voice customization and multilingual setup to reduce engineering dependency, and integrating advanced voice simulations (Duet simulations) with Duet Autopilot for faster testing and iteration. The post quantifies the impact with a ~37% reduction in time-to-go-live and more than halving of engineering hours per launch, enabling ADMs to launch voice agents with minimal agent engineer involvement.
  5. Conversational IVR from definition to business impact | DecagonThis post details the technical architecture and implementation of Decagon's Conversational IVR system. It outlines the five-step sequence: Automatic Speech Recognition (ASR), Natural Language Processing and Understanding (NLP/NLU), Dialogue Management, Text-to-Speech (TTS), and Continuous Learning. It highlights the importance of sub-second response times achieved through fine-tuned models and infrastructure co-location. The post also contrasts Conversational IVR with traditional IVR, emphasizing context-rich handoffs and the business benefits derived from higher containment rates, faster resolution, and improved customer satisfaction, citing specific cost reductions and ROI timelines. Implementation guidance includes starting with real customer utterances and choosing appropriate implementation paths.
  7. Beyond latency: The art of building a truly great voice agent | DecagonThis post details Decagon's approach to building great voice agents by emphasizing conversational design principles (efficiency, naturalness, recovery, empathy, user knowledge) and the importance of a synchronized voice profile (TTS characteristics and persona guidelines). It highlights the need for A/B testing and metric validation to discover the optimal voice profile and conversational design, moving beyond initial technical fundamentals.
  8. Bayesian VAD: Fixing turn detection in production voice pipelines | DecagonThis post introduces a Bayesian approach to Voice Activity Detection (VAD) and turn detection in voice AI pipelines. It addresses premature turn endings by first calibrating VAD probabilities using isotonic regression and then replacing frame-based heuristics with a Bayesian hazard statistic. This statistic accumulates evidence of silence and the likelihood of a true turn end, leading to a 42% reduction in broken turns and a 20% reduction in Pipeline WER.
  9. Outbound voice: Scaling proactive customer engagement | DecagonThis 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.
  10. Understanding enterprise AI answering services | DecagonThis post details the technical and compliance considerations for enterprise AI answering services, focusing on key functions like 24/7 availability, intelligent call handling, appointment scheduling, lead qualification, call routing, and CRM/system integration. It highlights the challenges of conversational complexity, action-taking capability, integration depth, and voice quality/latency. The post also outlines legal rules for inbound and outbound AI calls, emphasizing industry-specific compliance (HIPAA, GDPR). It introduces Decagon Voice as a platform that combines natural dialogue with authenticated action-taking, featuring real-time responsiveness, generation latency improvements, customizable voice profiles, cross-channel memory, smooth human escalations, SMS integration, and robust authentication methods.