We’ve released Conversational AI Engine v2.6 on April 22, 2026. This release focuses on improving agent control, conversation handling, and real-time state visibility, while expanding our TTS and multimodal capabilities.
What’s new in v2.6
New TTS provider: Deepgram
- Expands our TTS options with strong real-time performance and broader voice flexibility.
Inject instructions into an ongoing conversation
- A new capability allows clients to send custom text instructions directly into an agent’s active conversation flow.
- This enables scenarios such as implicit behavior tuning, client-side event triggering, and tighter voice + text coordination.
Cleaner MLLM turn handling
- Turn detection for MLLM has been refactored for better clarity and vendor-specific control (OpenAI Realtime API, Google Gemini Live).
- The result is more predictable turn-taking behavior across voice experiences.
Unified interruption control
- Agent interruption is now centrally managed via a single configuration.
- Supports both speech-based and keyword-based interruption, with clearer behavior when interruption is disabled.
Improvements
Clearer agent state signals
- Toolkits now expose explicit callbacks for when an agent is listening, thinking, or speaking—making it easier to build responsive UIs and debug live behavior.
More complete event payloads
- Agent name is now included in all NCS event payloads, improving observability and multi-agent scenarios.
For full technical details, please refer to the release notes