Nvidia used its annual GPU Technology Conference to detail a number of tools and technologies to help developers build AI capabilities into applications.

Among the announcements was availability of Nvidia Jarvis, an application framework for building conversational AI services. Its capabilities include highly accurate automatic speech recognition, real-time translation for multiple languages and text-to-speech capabilities, all designed to help developers create expressive conversational AI agents.

According to Nvidia, Jarvis is capable of real-time translation for five languages, at under 100ms latency per sentence, and expressive text-to-speech that delivers 30x higher throughput compared with the Tacotron2 neural network architecture. It is naturally intended for use with Nvidia GPUs, and Nvidia claimed it delivers 7x higher throughput than with CPUs.

Nvidia also announced the TAO framework, which stands for Train, Adopt and Optimise. This is a GUI-based, workflow-driven framework intended to simplify the creation of enterprise AI applications and services. It includes a diverse set of pre-trained models covering areas such as speech, vision, and natural language. By fine-tuning these pre-trained models, enterprises can produce domain-specific models in hours rather than months, Nvidia claimed, eliminating the need for large training runs and deep AI expertise. 

Nvidia Maxine SDKs, which have been created to enable developers to build virtual collaboration and content creation applications such as video conferencing and live streaming, are available now.

The release includes a Video Effects SDK for functions such as video noise removal and virtual backgrounds, an Augmented Reality SDK, and an Audio Effects SDK offering noise and room echo removal.

Nvidia also announced a number of updates. The latest release of Merlin, an open beta framework for the development of deep learning recommender systems, has added a new API and inference support to streamline the recommender workflow.

TensorRT 8.0 is the latest version of Nvidia’s high-performance deep learning inference SDK, which will be available in Q2 from the TensorRT page. It contains new features and optimisations to boost inference applications to run up to twice as fast with INT8 precision, with accuracy similar to FP32.

Nvidia DeepStream 6.0 is the latest release of the AI streaming analytics toolkit for building complex video analytics apps and services. It brings a new GUI and a suite of productivity tools that NVidia claimed will help build AI apps in days rather than weeks.