August 2026 Summaries

LiveKit has shifted billing for agent session minutes, recordings, WebRTC connections, and SIP connections from per-minute to per-second pricing, with a 10-second minimum charge, across all plans and enterprise contracts. The change requires no technical action from customers, is already active, and will appear on September invoices for August usage. LiveKit says the move is intended to better align costs with actual usage, particularly for short outbound calls that may reach voicemail, incorrect recipients, or end after only a few seconds, where minute-based billing can create disproportionate costs. By removing this mismatch, the company aims to make its platform more flexible for developers building specialized AI agents and broader outbound use cases.

Voice agents often fail to respond naturally to brief user replies such as “yes,” “no,” or “okay” because turn detection requires both an end-of-speech signal and a final speech-to-text transcript, either of which may be missing or delayed. Common causes include STT systems failing to finalize very short speech, overly aggressive endpointing that prematurely closes turns, permissive endpointing that creates long pauses, noise cancellation that mistakes speech for noise, and poor SIP or telephony audio quality. Recommended mitigations include enabling transcription timeouts to prompt users to repeat themselves, reviewing endpointing defaults and delays, correctly configuring or disabling noise cancellation when appropriate, improving telephony codecs and routing, and testing alternative STT providers or models. The discussion also distinguishes short answers from backchannels such as “uh-huh,” which can disrupt interruptions rather than cause silence, and recommends adaptive interruption handling. Diagnosis should rely on updated software, observability data, analysis of unusually long user turns, and systematic testing of transcription, audio-processing, turn-detection, and telephony settings.

LiveKit Cloud has introduced PII Redaction for Agent Observability, enabling developers to automatically remove sensitive information from agent call recordings, transcripts, logs, and traces before it is stored. The feature uses an LLM to identify personal data in conversational context, addressing challenges such as spoken spelling, corrections, and information spread across multiple turns, then replaces detected transcript content with labeled markers while dropping potentially sensitive trace and log fields. To prevent audio leakage, LiveKit removes entire turns containing PII and replaces them with a soft tone, while synchronized redacted transcripts preserve conversational context. LiveKit states that its runtime services use zero data retention, while redacted observability artifacts are securely retained for 30 days before automatic deletion. Available at no additional cost to Agent Observability customers using version 1.7.0 of the Python or Node.js Agents SDK, the feature supports 41 PII types across categories including names, contact details, financial data, locations, credentials, dates, and government identifiers, with configurable categories available through project observability settings.

LiveKit presents its SFU-based Rooms as a shared real-time pub/sub architecture for robotics systems involving robots, learned policies, human operators, recorders, planners, safety services, and cloud tools, allowing each participant to publish and selectively subscribe to media, state, and control tracks. The approach is intended to avoid indirect control paths and support low-latency coordination across distributed networks, with examples including human-in-the-loop recording, vision-language-model task orchestration, reward-model-driven policy escalation, and handoffs between classical planners and end-to-end learned policies. LiveKit Portal adds a standardized robot-operator interface with typed state, synchronized video and telemetry bundles, schemas, and command authority management so only the active operator can control the robot. For deployments requiring local latency, privacy, or high-bandwidth raw video, teams can run an on-premises SFU and bridge selected streams to LiveKit Cloud for remote operators, policies, dashboards, or recording without placing remote traffic in the local control loop. The company also describes a demonstration at Actuate 2026 featuring an SO-101 robotic arm that uses these components to respond to voice requests, select skills, assess execution, enable human takeover, and dispense candy.

LiveKit states that zero data retention (ZDR) is enabled by default for all LLM, speech-to-text, and text-to-speech models offered through its Inference platform, aiming to eliminate the provider-by-provider compliance reviews often required in regulated industries. It defines ZDR as using customer data only to process requests, with no retention, logging, training, aggregation, system improvement, or unauthorized sharing, and says it enforces these requirements through provider contracts, account configurations, workspace settings, and request-level parameters. The company reports that it verifies ZDR across its request entry points, model providers, and underlying GPU and cloud infrastructure before adding models and continuously reviews existing integrations. Some optional features necessarily retain data for their intended function: observability retains transcripts and audio for 30 days, Egress sends recordings to customer-controlled cloud storage while deleting interim files after upload, and voice cloning retains samples until 12 months after last use or until deletion by the customer.

LiveKit states that user data is not part of its business model and commits to never selling customer data or training models on it without explicit permission. The company emphasizes that its platform handles sensitive conversations, including health information and financial details, and says privacy protections apply by default rather than as optional features or paid tiers. It highlights zero data retention for every model in LiveKit Inference and says this standard is enforced across its services and third-party providers. LiveKit also previews future releases intended to improve conversational AI without relying on sensitive human training data, alongside planned updates on PCI DSS Level 1 and HIPAA compliance and PII redaction.

LiveKit has introduced expressive mode for its Agents framework to help voice AI adapt its tone, pacing, emotions, and nonverbal sounds to a user’s conversational context, addressing what it describes as a “feeling skills gap” that can reduce customer satisfaction in emotionally sensitive interactions. Enabled through a single configuration flag, the feature prompts an LLM to generate inline expressive markup, translates that markup into the dialect required by supported text-to-speech providers, cleans it from user-facing transcripts, and batches speech into larger chunks to preserve emotional consistency without meaningful added latency. It currently supports markup-capable LiveKit Inference models from Fish Audio, Inworld, Cartesia, and xAI, while allowing developers to selectively constrain delivery settings or customize the injected instructions. Expressive metadata is also made available to frontends through an lk.expression transcription attribute and a UI hook that converts provider-specific labels into normalized moods and colors for visualizers or indicators. LiveKit plans to add live delivery controls, broader provider support, and richer frontend expression data in future releases.

Agent join latency, the interval between dispatching an agent and its readiness in a LiveKit room, is distinct from conversational latency and can affect first impressions through ringing calls, silent clients, blank avatar screens, frontend timeouts, abandoned sessions, or rare token-expiration errors. Delays usually occur outside the server-to-worker handoff and commonly result from cold starts on LiveKit Cloud, unrealistic traffic bursts, large container images, blocking entrypoint code, avatar startup time, or insufficient warm capacity and poorly configured autoscaling in self-hosted deployments. Virtual avatars can extend the critical path because agents may wait for video to be published before speaking, while provider selection and avatar metrics can help assess performance. Self-hosted systems should maintain an appropriate pool of idle processes without exhausting CPU or memory, avoid burstable instances, and investigate warnings that no warmed process is available. Join time can be measured by correlating timestamps before dispatch and after the agent connects, while agents that never join may stem from mismatched agent names, token-based dispatch in pre-existing rooms, exhausted worker capacity, or blocked network traffic.