Product Updates - June 2026 : TurnV2 (Echo) Inference, NVIDIA Jetson Nano Support, and Advanced RTC Codecs

June 2026 Updates: TurnV2 (Echo) Inference, NVIDIA Jetson Nano Support, and Advanced RTC Codecs.

Welcome to the June edition of the VideoSDK Monthly Updates! This month, we have focused on making conversational AI interaction feel incredibly natural and expanding our native hardware capabilities.

We are introducing TurnV2 (Echo) low-latency models exclusively via our Inference Gateway, launching full native support for NVIDIA Jetson Nano in our C++ SDK, and rolling out selectable video codecs across our web and mobile platforms. Let's dive in!

AI Agents: Conversational Edge & TurnV2 (Echo)

While turn detection has always been a key component of VideoSDK, handling the fine-grained nuances of human conversation remains one of the hardest challenges in real-time AI.

Introducing TurnV2 (Echo) Low-Latency Models

We are launching TurnV2, a brand-new turn detector available exclusively via the VideoSDK Inference Gateway. It introduces two custom models:

echo-small : optimized for ultra-low latency response.
echo-large : optimized for high-accuracy turn classification.

Rather than simply checking for silence, the TurnV2 models classify speech into 4 complex conversational states, altering agent behavior dynamically:

Complete: The user has finished speaking. The agent responds immediately.
Incomplete: The user paused mid-sentence. The agent continues to listen.
Backchannel: The user makes brief acknowledging noises (like "uh-huh" or "yeah") while the agent is talking. The agent intelligently ignores this and keeps speaking.
Wait: The user explicitly commands the agent to stop. The agent interrupts its speech immediately and yields.

python

from videosdk.agents.inference import TurnV2

pipeline = Pipeline(
    stt=DeepgramSTT(),
    llm=GoogleLLM(),
    tts=CartesiaTTS(),
    vad=SileroVAD(),
    turn_detector=TurnV2.echo_large()  # or TurnV2.echo_small()
)

Unified Single-Line Imports

We have simplified developer workflows by introducing a unified component import style (v1.0.17). You can now import any provider plugin or inference gateway component from a single namespace, avoiding multi-line, verbose imports.

python

# Clean, unified imports for local plugins
from videosdk.agents.plugins import DeepgramSTT, GoogleLLM, CartesiaTTS

# Clean, unified imports for the Inference Gateway
from videosdk.agents.inference import SarvamAISTT, SarvamAITTS, NamoTurnDetectorV1

AWS Bedrock LLM Support

We have integrated support for AWS Bedrock via the ConverseStream API (v1.0.21). You can now leverage top-tier foundation models like Amazon Nova and Anthropic Claude within your pipeline, with enhanced fallback tool routing support.

python

from videosdk.agents.plugins import AWSBedrockLLM

pipeline = Pipeline(
    llm=AWSBedrockLLM(
        model="amazon.nova-lite-v1:0",
        region="us-east-1",
    )
)

Core RTC Upgrades: Codecs, DTX & Batched Signaling

We’ve introduced high-end streaming optimizations to our core JS, React, React Native, and iOS SDKs to keep your sessions lightweight and efficient.

Selectable Video Codecs

You can now specify preferred video codecs dynamically on initialization. The createCameraVideoTrack() method now accepts a `codec` parameter, supporting AV1, VP9, VP8, and H264.

DTX (Discontinuous Transmission)

Bandwidth optimization is now built-in. By enabling DTX across all audio profiles, the SDK automatically suspends transmission of audio data during silent moments, cutting bandwidth overhead significantly.

Batched Server Signaling

To reduce network traffic and client-side CPU consumption in larger meetings, the server now bundles multiple participant events into batched payloads of up to 5 events at a time. This behavior is enabled by default and can be customized via your meeting configuration.

PubSub Memory Safeguards

Available in React SDK v0.13.0, the new maxMessages and bufferMessages parameters allow you to cap or disable message retention on high-throughput topics, preventing unbounded memory leaks during fast data transmissions.

Embedded WebRTC: NVIDIA Jetson Nano Support

Our native C++ SDK continues to expand its target architectures, bringing WebRTC capabilities to rugged edge computers and IoT devices.

Jetson Nano Support (v0.0.1-beta.5): Official support has been added, making it straightforward to deploy real-time audio and video AI pipelines to Jetson Nano platforms.
VP8 Simulcast: Configure up to 3 outgoing video layers with customizable resolution, bitrate, and framerate properties before joining the meeting using setSimulcastLayers().
Native Data Channels: Exchange binary or text payloads of up to 15 KiB directly with peers on edge devices using the new sendData() method.

bash

curl -fsSL https://raw.githubusercontent.com/videosdk-live/videosdk-rtc-cpp-sdk/main/install.sh | sudo sh

Infrastructure & Telephony Connectors

Our server architecture received significant upgrades to support raw socket ingestion and simplified multi-tenant cloud deployments.

WebSocket Telephony Ingestion: Connect your custom telephone lines and gateways directly to VideoSDK using our standardized socket ingestion connectors. Explore the Architecture Guide.
Agent Cloud Registry: Easily deploy customized, containerized agents directly to the VideoSDK Cloud using our streamlined registry pipelines. Read the Deployment Guide.