📡 Adaptive bitrate streaming
Anam now dynamically adjusts video quality based on network conditions. When bandwidth drops, the stream adapts in real time to maintain smooth, uninterrupted video rather than freezing or dropping frames. When conditions improve, quality scales back up automatically. This is a significant improvement for users on mobile networks, VPNs, or connections with variable bandwidth.🔒 Zero Data Retention mode
Enterprise customers can now enable Zero Data Retention on any persona. When enabled, no session data — recordings, transcripts, or conversation logs — is stored after a session ends. This applies across the full pipeline including voice and LLM data.Toggle it on from persona settings in the Lab, or set it via the API. Learn more.Lab Changes
Improvements- System tools: Personas can now use built-in system tools.
change_languageswitches speech recognition to a different language mid-conversation, andskip_turnpauses the persona from responding when the user needs a moment to think. Enable them from the Tools tab in Build. - Tool validation: Auto-deduplication of tool names with clearer validation error messages.
- Share link management: Migrated share links to a 1-to-1 primary model with a simpler toggle interface.
- Fixed reasoning model responses getting stuck in “thinking…” state.
- Fixed soft-deleted knowledge folders not restoring on document upload.
- Fixed LiveKit session type classification for snake_case environment payloads.
Persona Changes
Improvements- Agora AV1 support: Agora integration now supports the AV1 video codec for better compression and quality at lower bitrates.
- Multi-agent LiveKit: Audio routing now works correctly in multi-agent LiveKit rooms with multiple Anam avatars.
- Fixed tool enum type validation.
February 27th, 2026
IntegrationsLabPersona
Four new integrations, Build page redesign & knowledge base overhaul
🔌 New integrations
Four new ways to use Anam avatars in your stack:PipecatThe
pipecat-anam package brings Anam avatars to Pipecat, the open-source framework for voice and multimodal AI agents. pip install pipecat-anam, add AnamVideoService to your pipeline, and you’re streaming. Use audio passthrough for full control over your own orchestration, or let Anam handle the pipeline end-to-end. GitHub repo.ElevenLabs server-side agentsPut a face on any agent you’ve built in ElevenLabs. Pass in your ElevenLabs agent ID and session token when starting a session, and Anam handles the rest, no changes to your existing ElevenLabs setup needed. Cookbook.VideoSDK
Anam is now officially supported on VideoSDK, a WebRTC platform similar to LiveKit. Built on top of the Python SDK.Framer
The Anam Avatar plugin is now on the Framer Marketplace. Drop an avatar into any Framer site without writing code.
📐 Metaxy: sample-level versioning for ML pipelines
We wrote up a deep dive on Metaxy, our open-source metadata versioning framework for multimodal data pipelines. It tracks partial data updates at the field level so teams only reprocess what actually changed. Works with orchestrators like Dagster, agnostic to compute (Ray, DuckDB, etc.). GitHub.Lab Changes
Improvements- Build page redesign: Everything lives in Build now. Avatars, Voices, LLMs, Tools, and Knowledge are tabs within a single page. Create custom avatars, clone voices, add LLMs, and upload knowledge files without leaving the page. Knowledge is a file drop on the Prompt tab: upload a document and it’s automatically turned into a RAG tool.
- Smart voice matching: One-shot avatars now auto-select a voice matching the avatar’s detected gender.
- Mobile improvements: Tables replaced with cards and lists. Bottom tab bar instead of hamburger menu. Long-press context menus on persona tiles. Touch-friendly tooltips.
- Knowledge base improvements: Non-blocking document deletion with pending state and rollback on error. PDF uploads restored. Stuck documents are auto-detected with retry from the UI.
- Fixed typo in thinking duration display.
- Fixed sticky hover states on touch devices.
Persona Changes
Improvements- Video stability: New TWCC-based frame-drop pacer with GCC congestion control. Smoother video on constrained or variable-bandwidth connections.
- Network connectivity: TURN over TLS for ICE, improving session establishment behind corporate firewalls and VPNs.
- Fixed ElevenLabs pronunciation issues with certain text patterns.
- Fixed text sanitization causing incorrect punctuation in TTS output.
- Fixed silent responses not being detected correctly.
SDK/API Changes
Improvements- Tool call event handlers:
onToolCallStarted,onToolCallCompleted, andonToolCallFailedhandlers for tracking tool execution on the client. - Documents accessed:
ToolCallCompletedPayloadnow includes adocumentsAccessedfield for Knowledge Base tool calls.
- Fixed duplicate tool call completion events.
🐍 Anam Python SDK
Anam now has a Python SDK. It handles WebRTC streaming, audio/video frame delivery, and session management.What’s in the box:- Media handling — The SDK manages WebRTC connections and signalling. Connect, and you get synchronized audio and video frames back.
- Multiple integration modes — Use the full pipeline (STT, LLM, TTS, Face) or bring your own TTS via audio passthrough.
- Live transcriptions — User speech and persona responses stream in as partial transcripts, useful for captions or logging conversations.
- Async-first — Built on Python’s async/await. Process media frames with async iterators or hook into events with decorators.
Lab Changes
Improvements- Visual refresh: Updated Lab UI with new brand styling, including new typography (Figtree), refreshed color tokens, and consistent component styles across all pages.
Persona Changes
Improvements- ICE recovery grace period: WebRTC sessions now survive brief network disconnections instead of terminating immediately. The engine detects ICE connection drops and holds the session open, allowing the client to reconnect without losing conversation state.
- Language configuration: You can now set a language code on your persona, ensuring the STT pipeline uses the correct language from session start.
- Voice generation options: Added configurable voice generation parameters for more control over TTS output.
- ElevenLabs streaming: Removed input buffering for ElevenLabs TTS, reducing time-to-first-audio for all sessions using ElevenLabs voices.
🎬 Session recordings
By default, every session is now recorded and saved for 30 days. Watch back any conversation in the Lab (lab.anam.ai/sessions) to see exactly how users interact with your personas, including the full video stream and conversation flow.Recordings and transcripts are also available via API. UseGET /v1/sessions/{id}/transcript to fetch the full conversation programmatically for analytics, QA, or archival. For privacy-sensitive applications, you can disable recording in your persona config.🎨 Two-pass avatar refinement
One-shot avatar creation now refines images in two passes. Upload an image, and the system generates an initial avatar, then refines it for better likeness and expression. Available to all users.Lab Changes
Improvements- Added
speechEnhancementLevel(0-1) tovoiceDetectionOptionsfor control over how aggressively background noise is filtered from user audio - Support for ephemeral tool IDs, so you can configure tools dynamically per session
- Added delete account and organization buttons
- Fixed terminology on tools tab
- Fixed RAG default parameters not being passed
- Fixed custom LLM default settings
Persona Changes
Improvements- Support for Gemini thinking/reasoning models
- The
speechEnhancementLevelparameter now passes through viavoiceDetectionOptions - Engine optimizations for lower latency under load
- Fixed GPT-5 tool calls returning errors
- Fixed audio frame padding that could cause playback issues
- Fixed repeated silence messages
- Fixed silence breaker not responding to typed messages
🎧 User Speech Enhancement
We’ve integrated ai-coustics as a preprocessing layer in our user audio pipeline. It enhances audio quality before it reaches speech detection, cleaning up background noise and improving signal clarity in real-world conditions. This reduces false transcriptions from ambient sounds and improves endpointing accuracy, especially in noisy environments like cafes, offices, or outdoor settings.🎛️ Configurable Persona Responsiveness
Control how quickly your persona responds with voiceDetectionOptions in the persona config:endOfSpeechSensitivity(0-1): How eager the persona is to jump in. 0 waits until it’s confident you’re done talking, 1 responds sooner.silenceBeforeSkipTurnSeconds: How long before the persona prompts a quiet user.silenceBeforeSessionEndSeconds: How long silence ends the session.silenceBeforeAutoEndTurnSeconds: How long a mid-sentence pause waits before the persona responds.
🧠 Reasoning Model Support
Added support for OpenAI reasoning models and custom Groq LLMs. Reasoning models can think through complex scenarios before responding, while Groq’s high-throughput infrastructure makes these typically-slower models respond with conversational latencies suitable for real-time interactions. Add your reasoning model in the lab: https://lab.anam.ai/llms.Persona Changes
Fixes- Fixed Knowledge Base (RAG) tool calling with proper default query parameters
- Fixed panic crashes when sessions error during startup
Lab Changes
Fixes- Fixed
Powered by Anamtext visibility when watermark removal is enabled - Updated API responses for GET/UPDATE persona endpoints
SDK/API Changes
Improvements- Introduced agent audio input streaming for BYO audio workflows, allowing you to integrate with arbitrary voice agents, eg ElevenLabs agents (see docs on how to integrate).
- Added WebRTC reasoning event handlers for reasoning model support
🎭 Introducing Cara 3: our most expressive model yet
The accumulation of over 6 months of research, Cara 3 is now available. This new model delivers significantly more expressive avatars featuring realistic eye movement, more dynamic head motion, smoother transitions in and out of idling, and improved lip sync.You can opt-in to the new model in your persona config usingavatarModel: 'cara-3' or by selecting it in the Lab UI. Note that all new custom avatars will use Cara 3 exclusively, while existing personas will continue to use the Cara 2 model by default unless explicitly updated.🛡️ SOC-2 Type II compliance
Anam has achieved SOC-2 Type II compliance. This milestone validates that our security, availability, and data protection controls have been independently audited and proven over time.For customers building across learning, enablement, or live production use cases, this provides formal assurance regarding how we handle security, access, and reliability.Visit the Trust Center
🔌 Integrations
Model Context Protocol (MCP) serverManage your personas and avatars directly within Claude Desktop, Cursor, and other MCP-compatible clients. Use your favorite LLM-assisted tools to interact with the Anam API.Anam x ElevenLabs agents
Turn any ElevenLabs conversational AI agent into a visual avatar using Anam’s audio passthrough.
Watch the demo
Lab Changes
Improvements- UI overhaul: A redesigned Homepage and Build page make persona creation more intuitive. You can now preview voices/avatars without starting a chat and create custom assets directly within the Build flow. Sidebar and Pricing pages have also been refreshed.
- Performance: Implemented Tanstack caching to significantly improve Lab responsiveness.
- Fixed a bug where client tool events were not appearing in the Build page chat.
- Resolved an issue where tool calls and RAG were not passing parameters correctly.
Persona Changes
Improvements- More voices: Added ~100 new Cartesia voices (Sonic-3) and ~180 new ElevenLabs voices (Flash v2.5), covering languages and accents from all over the world.
- New default LLM:
kimi-k2-instruct-0905is now available. This SOTA open-source model offers high intelligence and excellent conversational abilities. (Note: Standardkimi-k2remains recommended for heavy tool-use scenarios). - Configurable greetings: Added
skip_greetingparameter, allowing you to configure whether the persona initiates the conversation or waits for the user. - Latency reductions:
- STT optimization: We are now self-hosting Deepgram for Speech-to-Text, resulting in a ~30ms (p50) and ~170ms (p90) latency improvement.
- Frame buffering: Optimized output frame buffer, shaving off an additional ~40ms of latency per response.
- Corrected header handling to ensure reliable data center failover.
- Fixed a visual artifact where Cara 3 video frames occasionally displayed random noise.
- Resolved a freeze-frame issue affecting ~1% of sessions (Incident Report).
SDK/API Changes
Improvements- API gateway guide: added documentation and an example repository for routing Anam SDK traffic through your own API Gateway server. View on GitHub.
🎥 Livekit out of Beta and new latency record
LiveKit integration is now generally available: drop Anam’s expressive real-time avatars into any LiveKit Agents app so your AI can join LiveKit rooms as synchronised voice + video participants.It turns voice-only agents into face-and-voice experiences for calls, livestreams, and collaborative WebRTC spaces, with LiveKit handling infra and Anam handling the human layer. Docs
⚡ Record-breaking latency: 330 ms decrease in latency for all customers
Server-side optimisations cuts average end-to-end latency by 330 ms for all customers, thanks to cumulative engine optimisations across transcription, frame generation, and frame writing, plus upgraded Deepgram Flux endpointing for faster, best in class turn-taking without regressions in voice quality or TTS.Lab Changes
Improvements • Overhaul to avatar video upload and management system• Upgraded default Cartesia voices to Sonic 3• Standardised voice model selection across the platformFixes • Enhanced share link management capabilities• Corrected LiveKit persona type identification logicPersona Changes
Improvements • Server-side optimisations to our frame buffering to reduce latency of responses by ~250ms for all personas.Fixes • Changed timeout behavior to never time out based on heartbeats; only time out when websocket is disconnected for 10 seconds or more.• Fixed intermittent issue where persona stopped responding• Set pix_fmt for video output, moving from yuvj420p (JPEG) to yuv420 color space to avoid incorrect encoding/output.• Added timeout in our silence breaking logic to prevent hangs.🚀 Introducing Anam Agents
Build and deploy AI agents in Anam that can engage alongside you.With Anam Agents, your Personas can now interact with your applications, access your knowledge, and trigger workflows directly through natural conversation. This marks Anam’s evolution from conversational Personas to agentic Personas that think, decide, and execute.Knowledge Tools
Give your Personas access to your company’s knowledge. Upload docs to the Lab, and they’ll use semantic retrieval to integrate the right info.Docs for Knowledge Base
Client Tools
Personas can control your interface in real time—open checkout, display modals, navigate UI, and update state by voice.Docs for Client Tools
Webhook Tools
Connect your Personas to external APIs and services. Create tickets, fetch status, update records, or fetch live data.Docs for Webhook Tools
Intelligent Tool Selection
Each Persona’s LLM chooses tools based on intent—not scripts.You can create/manage tools on the Tools page in the Lab and attach them to any Persona from Build.Anam Agents are available in beta for all users: https://lab.anam.ai/loginLab Changes
Improvements- Cartesia Sonic-3 voices: the most expressive TTS model.
- Voice modal expanded: 50+ languages, voice samples, Cartesia TTS now default.
- Session reports work for custom LLMs.
- Prevented auto-logout when switching contexts.
- Fixed race conditions in cookie handling.
- Resolved legacy session token issues.
- Removed problematic voices.
- Corrected player/stream aspect ratios on mobile.
Persona Changes
Improvements- Deepgram Flux support for turn-taking (Deepgram Flux Details)
- Server-side optimization: reduced GIL contention and latency, faster connections.
- Bug-fix for dangling LiveKit connections.
Research
Improvements🛡️ Anam is now HIPAA compliant
A big milestone for our customers and partners. Anam now meets HIPAA requirements for handling protected health information.Learn more at the Anam Trust CenterLab Changes
Improvements- Enhanced voice selection: search by use case/conversational style, 50+ languages.
- Product tour update.
- Streamlined One-Shot avatar creation.
- Auto-generated Persona names based on selected avatar.
- Session start now 1.1s faster.
- Share links: fixed extra concurrency slot usage.
Persona Changes
Improvements- Improved TTS pronunciation via smarter text chunking.
- Traceability and monitoring for session IDs.
- Increased internal audio sampling rate to 24kHz.
- Increased max websocket size to 16Mb.
- Concurrency calculation now only considers sessions from last 2 hours.
- Less freezing for slower LLMs.
📊 Session Analytics
Once a conversation ends, how do you review what happened? To help you understand and improve your Persona’s performance, we’re launching Session Analytics in the Lab. Now you can access a detailed report for every conversation, complete with a full transcript, performance metrics, and AI-powered analysis.- Full Conversation Transcripts. Review every turn of a conversation with a complete, time-stamped transcript. See what the user said and how your Persona responded, making it easy to diagnose issues and identify successful interaction patterns.
- Detailed Analytics & Timeline. Alongside the transcript, a new Analytics tab provides key metrics grouped into “Transcript Metrics” (word count, turns) and “Processing Metrics” (e.g., LLM latency). A visual timeline charts the entire conversation, showing who spoke when and highlighting any technical warnings.
- AI-Powered Insights. For a deeper analysis, you can generate an AI-powered summary and review key insights. This feature, currently powered by gpt-5-mini, evaluates the conversation for highlights, adherence to the system prompt, and user interruption rates.
Lab Changes
Improvements- Improved Voice Discovery: The Voices page has been updated to be more searchable, allowing you to preview voices with a single click, and view new details like gender, TTS-model and language.
- Fixed share-link session bug: Fixed bug of share-link sessions taking an extra concurrency slot.
Persona Changes
Improvements- Small improvement to connection time: Tweaks to how we perform webrtc signalling which allows for slightly faster connection times (~900ms faster for p95 connection time).
- Improvement to output audio quality for poor connections: Enabled Opus in-band FEC to improve audio quality under packet loss.
- Small reduction in network latency: Optimisations have been made to our outbound media streams to reduce A/V jitter (and hence jitter buffer delay). Expected latency improvement is modest (<50ms).
- Fix for livekit sessions with slow TTS audio: Stabilizes LiveKit streaming by pacing output and duplicating frames during slowdowns to prevent underflow.
⚡ Intelligent LLM Routing for Faster Responses
The performance of LLM endpoints can be highly variable, with time-to-first-token latencies sometimes fluctuating by as much as 500ms from one day to the next depending on regional load. To solve this and ensure your personas respond as quickly and reliably as possible, we’ve rolled out a new intelligent routing system for LLM requests. This is active for both our turnkey customers and for customers using their own server-side Custom LLMs if they deploy multiple endpoints.This new system constantly monitors the health and performance of all configured LLM endpoints by sending lightweight probes at regular intervals. Using a time-aware moving average, it builds a real-time picture of network latency and processing speed for each endpoint. When a request is made, the system uses this data to calculate the optimal route, automatically shedding load from any overloaded or slow endpoints within a region.Lab Changes
Improvements- Generate one-shot avatars from text prompts: You can now generate one-shot avatars from text prompts within the lab, powered by Gemini’s new Nano Banana model. The one-shot creation flow has been redesigned for speed and ease-of-use, and is now available to all plans. Image upload and webcam avatars remain exclusive to Pro and Enterprise.
- Improved management of published embed widgets: Published embed widgets can now be configured and monitored from the lab at https://lab.anam.ai/personas/published.
Persona Changes
Improvements- Automatic failover to backup data centres: To ensure maximum uptime and reliability for our personas, we’ve implemented automatic failover to backup data centres.
- Prevent session crash on long user speech: Previously, unbroken user speech exceeding 30 seconds would trigger a transcription error and crash the session. We now automatically truncate continuous speech to 30 seconds, preventing sessions from failing in these rare cases.
- Allow configurable session lengths of up to 2 hours for Enterprise plans: We had a bug where sessions had a max timeout of 30 mins instead of 2 hours for enterprise plans. This has now been fixed.
- Resolved slow connection times caused by incorrect database region selection: An undocumented issue with our database provider led to incorrect region selection for our databases. Simply refreshing our credentials resolved the problem, resulting in a ~1s improvement in median connection times and ~3s faster p95 times. While our provider works on a permanent fix, we’re actively monitoring for any recurrence.
🔌 Embed Widget
Embed personas directly into your website with our new widget. Within the lab’s build page click Publish then generate your unique html snippet. This snippet will work in most common website builders, eg Wordpress.org or SquareSpace.For added security we recommend adding a whitelist with your domain url. This will lock down the persona to only work on your website. You can also cap the number of sessions or give the widget an expiration period.Lab Changes
Improvements- ONE-SHOT avatars available via API: Professional and Enterprise accounts can now create one-shot avatars via API. Docs here.
- Spend caps: It’s now possible to set a spend cap on your account. Available in profile settings.
Persona Changes
Fixes- Prevent Cartesia from timing out when using slow custom LLMs: We’ve added a safeguard to prevent Cartesia contexts from unexpectedly closing during pauses in text streaming. With slower llms or if there’s a break or slow-down in text being sent, your connection will now stay alive, ensuring smoother, uninterrupted interactions.