Changelog

​

New Features

• Magic Canvas image support: The image component now renders images from your Knowledge Base or an allowlisted website during a conversation. Learn more
• Auto-start recordings: Set auto_start_recording on create-conversation and Tavus starts the recording about a second after the PAL joins, with no client-side call to startRecording(). Requires recording_storage and a Tavus-hosted room. Learn more

​

Changes

• New default LLM - tavus-gemma-4: A faster, smarter, tuned version of tavus-gemma-4 is now the recommended default. It makes for the snappiest conversations while staying strong on intelligence and tool use. Learn more
• Deprecated LLM: tavus-gemma-4-thinking is deprecated in favor of the new tuned tavus-gemma-4.

​

New Features

• EU AI Act controls: New AI disclosure controls on the PAL - disclosure_type, verbal_disclosure, and visual_disclosure - plus a per-conversation policy: "eu" that auto-applies the disclosure and switches Raven-1 to limited emotion recognition when PAL fields are left on auto. Learn more
• Zoom support: PALs can now join Zoom calls in addition to Google Meet. Invite the PAL’s @tavusinvite.com address to a calendar event with a Zoom link and it joins on its own. Learn more

​

New Features

• Expanded Tavus-Hosted LLM Selection: Added support for new Tavus-hosted LLMs: tavus-gemma-4, tavus-gpt-5.6-sol, and tavus-gpt-5.6-terra. Learn more

​

Changes

• Deprecated LLMs: The following Tavus-hosted LLMs are now deprecated: tavus-glm-4.7, tavus-gpt-oss, tavus-claude-haiku-4.5, tavus-gpt-5.2, and tavus-gemini-3-flash.

​

New Features

• Zero Data Retention: Account-level and programmatic support for Zero Data Retention is now available, enabling customers to opt out of data persistence across Tavus services.

​

Enhancements

• Performance: Improved LiveKit pipeline performance by 15%!

​

New Features

• PAL Maker: Describe a PAL in plain language and have it configured end to end (system prompt, face, objectives, guardrails, tools, knowledge base, and advanced settings), then talk to it in the same session. Removes the need to configure each layer by hand. Learn more
• Starter Templates: Fork a complete, live experience into your account with no code (Interviewer, SDR, Medical Intake, Trivia Host), or open the GitHub repo behind it. Learn more
• Meetings (Google Meet): Send a PAL into a Google Meet call. It joins on its own past the admit prompt, sees and hears the room, and can run the meeting. Invite it via calendar invite or pass a meeting URL when creating a conversation. Learn more
• Magic Canvas: A shared interactive surface rendered on top of the video stream. The PAL surfaces the right component at the right moment. Six components at launch: question, text, input, calendar, chart, and alert. Learn more
• Presentation Mode: Upload a deck or source material, Tavus indexes it, and the PAL presents from it, takes questions, and pulls the right slide on demand. Supports two modes: Walk the deck (end to end) and On-demand. Learn more
• Embeddable Widgets: Drop a PAL onto any site with a single line of code, scoped and auto-provisioned to your account, with no SDK. Place it in any corner or inline, and customize branding to match your site. Learn more
• Landing Pages: Deploy a PAL on a ready-made hosted landing page, no setup required. Learn more
• Internet Search: Give a PAL real-time web search during a conversation. Attach the internet_search skill - pure on/off, no configuration. Learn more

​

Enhancements

• Tool Calling: Tools can now call any third-party API (including GraphQL) with custom auth and request bodies, no middleware. New on-resolve modes (fire and forget, pull into context, generate a response) control how results enter the conversation, and tools are reusable across PALs. Learn more

​

Enhancements

• Meet Rivian, Tiffany, and Brian: Our 3 new stock replicas!
• Improved Image to Replica uploading flow: Get more detailed and accurate errors before training submission.

​

New Features

• Guardrails as first-class primitives: Guardrails are now standalone resources that can be created, edited, and reused across multiple PALs. Compose individual guardrails into sets via guardrail_tags. Learn more
• Choose your delivery channel: Guardrail violations can now be delivered via Interaction Events (in-call app-message) or Webhooks (server-side callback_url) - or both.
• guardrail_uuid on violations: Triggered events and callbacks now include guardrail_uuid, so you can identify exactly which individual guardrail was violated.

​

New Features

• Azure TTS support: Azure is now available as a TTS engine, expanding language coverage for multilingual and localized voice output. Learn more

​

New Features

• Individually addressable guardrails: Guardrails are now first-class resources. Create, attach, edit, and delete each guardrail independently via the Guardrails API. Bundle guardrails by tag and reference them on a PAL via guardrail_ids or guardrail_tags. Learn more
• Node.js plugin support for LiveKit integration: The LiveKit Agents integration now supports Node.js via the @livekit/agents-plugin-tavus plugin, in addition to Python. Install with npm install @livekit/agents @livekit/agents-plugin-tavus. Learn more
• Real-time event timestamps: Every interaction event (conversation.utterance, conversation.started_speaking, etc.) now carries a timestamp field delivered in real time on the respective Interaction event. Learn more
• Transcript utterance timestamps: End-of-call transcripts now carry per-turn timing. The application.transcription_ready webhook (and the same payload nested under events in the verbose GET /conversations/{id}?verbose=true response) now includes timestamp (Unix epoch float, seconds - same field name as live interaction events), seconds_from_start, duration (seconds, float - same field name as conversation.stopped_speaking.duration), and inference_id (on assistant turns) on each transcript entry.

​

Enhancements

• New, and more detailed, error messages when a face fails: Learn more

​

New Features

• AI Image Fixer API support: Create Face now accepts an auto_fix_training_image property. Set it to true to use Tavus’s AI Image Fixer to instantly fix any uploaded image to fit our requirements, eliminating the need for editing or recapturing photos. Learn more

​

New Features

• AI Image Fixer: Instantly fix any uploaded image to fit our requirements, eliminating the need for editing or recapturing photos.

​

New Features

• Frame Checker (Video to Face): Get instant feedback, before recording, on whether your camera setup meets our requirements.

​

Enhancements

• Less footage required (Video to Face): We now only require 1 minute of video, down from the 2 minutes previously needed.
• Simpler in-portal recording flow (Video to Face): A streamlined recording experience in the portal to help you capture high-quality recordings.

​

New Features

• Image to Face: Build a face from a single still. Drop in a photo, illustration, or brand mascot.

​

New Features

• Voice Activity Detector improvements: Resulting in a smoother conversational experience in noisy environments. This is automatically rolled out to all users.
• Expanded Recording Storage Support: Conversation recordings can now be delivered to Google Cloud Storage (GCP) and Azure Blob Storage, in addition to AWS S3. Learn more

​

Enhancements

• Tavus Components Library Updates: Improved audio-video sync, plus new chat components and closed captions with streaming support. Learn more

​

New Features

• Wake Phrase: PALs can now stay silent until they hear a specific phrase, similar to how voice assistants like Siri or Alexa work. Configure it via the wake_phrase parameter in the Conversational Flow layer. The PAL still hears everything that is said and responds with full conversation history once the wake phrase is detected. Learn more

​

New Features

• Idle Engagement: Faces can now proactively re-engage by speaking to the user after a period where the user is silent. Eagerness of this feature can be configured via the idle_engagement parameter in the Conversational Flow layer. Learn more

​

New Features

• Speaking Events: Two new events - conversation.started_speaking and conversation.stopped_speaking - fire for both the PAL and the user with a role field ("pal" or "user") identifying the speaker. Tavus also sends legacy duplicate events with "replica" instead of "pal". conversation.stopped_speaking includes an interrupted boolean and a duration field (in seconds). Learn more
• Conversation Diagnostics: A new diagnostics surface for inspecting what happened in a conversation - including packet loss, network connection, FPS, and more - designed to make debugging significantly faster. Click on any conversation in the PAL Maker to access its diagnostics page.

​

Enhancements

• Non-Interruptible Custom Greetings: Custom greetings now finish entirely before users can interrupt speech. Previously, participants could talk over a custom_greeting; now the PAL completes the greeting before it begins listening. Learn more
• Improved Turn-Taking Latency: Significant TTS optimizations reduce turn-taking latency, resulting in faster and more natural back-and-forth during conversations.

​

New Features

• Streaming Utterance Event: A new conversation.utterance.streaming event progressively reports what has been said during a conversation turn for both face and user utterances. Use it to power closed captioning and build accurate transcripts - especially when a user interrupts the face, since the streaming event reflects only the words actually spoken rather than the full LLM response. Learn more
• Pronunciation Dictionaries: Define custom pronunciation rules so your PAL says brand names, technical terms, acronyms, and foreign words exactly right. Choose between simple alias substitution (e.g., “Tavus” → “TAH-vus”) or precise IPA phonetic notation. Create a dictionary once and attach it to a PAL via the TTS layer - any updates automatically propagate to all linked PALs with zero extra latency at conversation time. Learn more

​

New Features

• Voice Isolation: Filter background noise from participant audio to improve conversation quality. Configure it via the voice_isolation parameter in the Conversational Flow layer. Learn more

​

Changes

• Chat Interrupt History: PALs now know when they have been interrupted. This allows the PAL to pick back up where it left off, and also improves objectives adherence.

​

New Features

• Expanded ASR Model Selection: You can now choose from five specialized speech-to-text engines via the stt_engine parameter. New models include tavus-parakeet, tavus-soniox, tavus-whisper, and tavus-deepgram-medical. Use tavus-auto to automatically route to the best model for each conversation. Learn more

​

Enhancements

• 30% Faster Phoenix-4 Boot Time: Phoenix-4 conversations now boot 30% faster, significantly reducing the time from conversation creation to readiness.

​

Changes

• conversation.replica_interrupted Event Removed: The conversation.replica_interrupted application message has been removed from interaction events. This event was deprecated in a previous backend update. Use conversation.replica.stopped_speaking with the interrupted: true property to detect interruptions instead.
• duration and interrupted Fields on Face Stopped Speaking: The conversation.replica.stopped_speaking event now includes a duration field (how long the PAL spoke in seconds) and an interrupted field (true/false) indicating whether the PAL was interrupted by the user. Learn more

​

New Features

• Event Ordering and Turn Tracking: All server-broadcasted interaction events now include seq and turn_idx fields. seq is a globally monotonic sequence number for ordering events that may arrive out of order, and turn_idx groups related events from the same conversation turn. Learn more

​

Enhancements

• 30% Faster Phoenix-4 Boot Time: Phoenix-4 conversations now boot 30% faster, significantly reducing the time from conversation creation to readiness.

​

Enhancements

• EU ElevenLabs BYOK Support: Customers can now bring their own ElevenLabs API key from EU-region accounts.

​

Enhancements

• Improved Knowledge Base Retrieval: Optimized underlying infrastructure to improve utterance to utterance response times, particularly when rag_search_quality is set to quality.

​

New Features

• Expanded Tavus-Hosted LLM Selection: Added new Tavus-hosted LLM options including models from Gemini, Claude, and GPT families. tavus-gpt-oss is recommended as the default. Legacy models tavus-gpt-4.1, tavus-gpt-4o, and tavus-gpt-4o-mini are now deprecated. Learn more →
• Visual RAG: CVI now supports visual retrieval-augmented generation. Upload custom image explanations that are matched and queried via vision embeddings, giving your PAL richer visual context during conversations.

​

Changes

• PALcontextField Deprecated: The context field has been deprecated in favor of a unified system_prompt field. Existing context values have been automatically merged into system prompts. The API remains backward compatible, but we recommend using only system_prompt going forward.

​

New Features

• Raven-1 Perception Model: Introduced Raven-1, a multimodal perception model with audio emotion analysis and enhanced visual awareness. Raven-1 captures user emotion from audio in real time (sub-100ms audio perception latency), enabling PALs to respond with greater emotional intelligence. The model is now the default for all new PALs. Enable it by setting perception_model_name in your PAL configuration. Learn more →
• Private Rooms: Require authentication to join conversations for enhanced security. When enabled, we return a JWT meeting token that users must include when entering the room. Learn more

​

Enhancements

• Upgraded Transcription Engine: Upgraded transcription engine with 3x improvements in word error rates (WER).

​

New Features

• Website Crawling for Knowledge Base: You can now enable link crawling when creating knowledge base documents. Configure crawl depth and max_pages to automatically discover and ingest content from linked pages. Additionally, existing crawled documents can now be recrawled to keep knowledge base content up to date.

​

Changes

• PlayHT TTS Removed: PlayHT has been fully removed as a supported TTS engine. All PALs previously using PlayHT should migrate to Cartesia or ElevenLabs.

​

New Features

• Hard Delete for Conversations: Conversations can now be permanently deleted via the API using the hard=true query parameter. Use this for GDPR compliance or data cleanup workflows.

​

Enhancements

• Default TTS Model Updated to Sonic-3: The default text-to-speech model has been updated to Sonic-3 across all new PALs, delivering improved voice quality and naturalness.
• LiveKit Connection Stability: Extensive reliability improvements to the LiveKit-based transport layer, including fixes for connection timeouts, track publishing hangs, event loop starvation, and ping timeout issues.

​

Changes

• Default LLM Migrated to tavus-gpt-oss: The default LLM for all new PALs is now tavus-gpt-oss. All remaining tavus-llama-4 PALs have been automatically migrated. Legacy Tavus-Llama model references have been removed.

​

New Features

• LLM Temperature & Top-P Parameters: You can now configure temperature and top_p parameters for both Tavus-hosted LLMs and custom LLMs via the extra_body field in your PAL’s LLM configuration. Learn more →

​

Enhancements

• Text Echo Language Accuracy: Text echoes now correctly use the input language for conversion, improving accuracy in multilingual conversations.

​

New Features

• Test Mode for Conversations: You can now start conversations in test mode, where the PAL does not join. Validate your setup, integrations, and conversational flows without incurring costs or using concurrency slots. Set test_mode: true when creating a conversation. Learn more →

​

Enhancements

• Fuzzy Search for PALs: Search now supports fuzzy matching for PALs, allowing users to find results based on partial matches of UUIDs or names.

​

New Features

• Memories: CVI now remembers context across conversations. Every conversation builds on the last with full context and time/date awareness, enabling use cases like adaptive tutoring, mentorship, and recurring consultations. Learn more →
• Knowledge Base (RAG): Bring your own data to conversations instantly. Upload documents or links and get grounded answers with ~30ms retrieval latency. Power AI recruiters, support agents, travel guides, and more with domain-specific knowledge. Learn more →
• Objectives & Guardrails: Define clear goals, branching logic, and measurable outcomes for your PALs while keeping conversations safe, compliant, and on-brand. Ideal for complex workflows and regulated industries. Learn more →
• PAL Builder: A guided creation flow in the PAL Maker to shape AI PALs with goals, behaviors, and style - then test or launch within minutes.

​

New Features

• Events Console: A new events console in the PAL Maker lets you monitor everything happening during a conversation in real time - from message flows to system activity.
• Conversation Transcripts & Perception Analysis: View full conversation details directly in the PAL Maker, including transcripts with speaker roles and perception analysis showing how your AI PAL sees, hears, and responds.

​

New Features

• PAL Layer Controls: Enable or disable layers like Sparrow directly within a PAL and adjust sensitivity settings in real time from the PAL Maker side panel.
• PAL editing in PAL Maker: We’ve added new editing capabilities to help you refine your PALs more efficiently. You can now update system prompt, context, and layers directly in our PAL Maker, plus duplicate existing PALs to quickly create variations or use them as starting points for new projects. Find these new features in your PAL Library at maker.tavus.io/dev.

​

Enhancements

• Interaction Events Playground Improvements: Major updates to the Interaction Events Playground including correct properties.context format and append vs overwrite toggle.

​

New Features

• Multilingual Settings in PAL Maker: You can now specify the language of a conversation directly in the PAL Maker, including a new multilingual option for dynamic, real-world interactions.

​

New Features

• Llama 4 Support: Your PAL just got even smarter, thanks to Meta’s Llama 4 model 🧠 You can start using Llama 4 by specifying tavus-llama-4 for the LLM model value when creating a new PAL or updating an existing one. Click here to learn more!

​

New Features

• React Component Library: Developers can build with Tavus even faster now with our pre-defined components 🚀 Click here to learn more!

​

New Features

• Multilingual Conversation Support: CVI now supports dynamic multilingual conversations through automatic language detection. Set properties.language to “multilingual” and CVI will automatically detect the user’s spoken language and respond in the same language using ASR technology.
• Audio-Only Mode: CVI now supports audio-only conversations with advanced perception (powered by Raven) and intelligent turn-taking (powered by Sparrow-1). Set audio_only=true in your create conversation request to enable streamlined voice-first interactions.

​

Enhancements

• Fixed CVI responsiveness issue: Resolved an issue where CVI would occasionally ignore very brief user utterances. All user inputs, regardless of length, now receive consistent responses.
• Expanded tavus-llama-4 context window: Increased maximum context window to 32,000 tokens. For optimal performance and response times, we recommend staying under 25,000 tokens.

​

Enhancements

• Reduced conversation boot time by 58% (p50).

​

Changes

• Added a new recording requirement to Training from a video : Start the talking segment with a big smile.

​

Enhancements

• Added echo and respond events to conversational context.

​

Enhancements

• Major Phoenix 3 Enhancements for CVI:
  • Increased frame rate from 27fps to 32fps, significantly boosting smoothness.
  • Reduced Phoenix step’s warm boot time by 60% (from 5s to 2s).
  • Lipsync accuracy improved by ~22% based on AVSR metric.
  • Resolved blurriness and choppiness at conversation start.
  • Enhanced listening mode with more natural micro expressions (eyebrow movements, subtle gestures).
  • Greenscreen mode speed boosted by an additional ~1.5fps.
• Enhanced CVI Audio Quality: Audio clicks significantly attenuated, providing clearer conversational audio.
• Phoenix 3 Visual Artifacts Fix: Resolved visual artifacts in 4K videos on Apple devices, eliminating black spot artifacts in thumbnails.

​

New Features

• Launched LiveKit Integration : With Tavus video agents now integrated into LiveKit, you can add humanlike video responses to your voice agents in seconds.
• PAL API : Enabled patch updates to PALs.

​

Enhancements

• Resolved TTS (Cartesia) stability issues and addressed hallucination.
• Phoenix 3 Improvements:
  • Fixed blinking/jumping issues and black spots in videos.
  • FPS optimization to resolve static and audio crackling.

​

Enhancements

• Face API:
  • Enhanced Error Messaging for Training Videos.
  • Optimized Auto QA for Training Videos.