Speechmatics Supercharges Machine Learning Capabilities with NVIDIA AI

Deep tech firm bursts through processing ceiling as it scales globally

Speechmatics, the world’s leading speech-to-text API scaleup, has announced it is using NVIDIA accelerated computing to supercharge its machine learning capabilities. Speechmatics’ transition to self-supervised learning has increased the need for processing power. Speechmatics has now deployed NVIDIA market-leading artificial intelligence (AI) systems and software stack to train larger models and take its technology ecosystem to the next level.

Speechmatics is one of the first companies to upgrade to the NVIDIA DGX™ H100 system featuring eight new H100 Tensor Core graphics processing units (GPUs), which provides a step change in the speed and efficiency of training the next generation of neural networks. With speech-to-text becoming an integral feature for multiple industries, including contact centres, media and event captioning, media monitoring, meeting platform solutions, and video distribution platforms, an efficient and scalable solution is critical. NVIDIA leads the market in accelerated computing, with a recent report highlighting the company’s 80% share in the industry, making the NVIDIA DGX H100 the obvious choice for Speechmatics.

Speechmatics’ speech-to-text API is built using self-supervised machine learning which requires huge datasets for training and large amounts of computation. Staying at the forefront of innovation in research also requires delivering on architecture and infrastructure for training larger AI models. The architecture of the NVIDIA DGX H100 system makes it optimal for running the next generation of transformer models – the backbone of self-supervised learning.

The Tensor Cores in the NVIDIA H100 GPU offer double the speed of matrix multiplication – the building blocks of neural networks – while the large memory capacity of the DGX H100 system allows Speechmatics to use models twice the size of the ones previously. Additionally, the advanced networking capabilities, using the NVIDIA Quantum InfiniBand platform and – NVIDIA NVLink technology, can deliver 9x higher bandwidth versus previous generations.

“Natural language processing is powering innovative, informative services that can grow customer success and revenue for enterprises,” said Tony Paikeday, Senior Director of AI systems at NVIDIA. “As a fully integrated AI platform, NVIDIA DGX H100 systems provide a powerful foundation for training speech AI models to further accelerate Speechmatics.”

Will Williams, VP Machine Learning at Speechmatics added: “Speechmatics aims to be the first company in the world to reach perfect-level speech-to-text performance. Its next-generation models will be built on and optimised for NVIDIA technology with larger modelling capacity being at the core of their R&D roadmap. Next-generation modelling capability will always need to be backed with next-generation hardware capability. We can expect another step change in automated speech recognition performance purely by harnessing the latest NVIDIA GPUs."