Achronix unveils FPGA-powered automatic speech recognition solution

1 min read

Achronix, in partnership with, has developed an accelerated automatic speech recognition (ASR) solution based on the company’s Speedster7t FPGA.

This solution converts spoken language to text in over 1,000 concurrent real-time streams with high accuracy and fast response times and is said to be able to deliver up to a 20 times improvement in performance over competing solutions.

The solution is powered by a VectorPath accelerator card featuring a Speedster7t FPGA running’s Achronix-FPGA-optimised ASR IP - delivering real-time, ultra-low latency speech-to-text capabilities.

A single card in a server can replace up to 20 CPU-only-based servers or 15 GPU cards. The AI model can also be customised to trade off accuracy versus performance when support for 1,000 concurrent streams is unnecessary.

According to Achronix, it is poised to disrupt the ASR landscape with its exceptional word-error rate and 99th percentile latency of 54 ms end-to-end, including pre- and post-processing plus data movement back and forth to the CPU. In addition, the solution can be customised or retrained with vertical-specific or custom data sets in standard machine learning (ML) frameworks.

"One of the key aspects of the accelerated ASR solution built on Achronix Speedster7t FPGAs is its ability to reduce both OpEx and CapEx while maintaining top-tier performance significantly," said Bill Jenkins, the Director of AI Product Marketing at Achronix. "This solution, powered by a Speedster7t FPGA, can reduce costs by up to 90% compared to traditional CPU/GPU-based server solutions, whether enterprise or in the cloud. This capability translates to tangible business savings while providing exceptional real-time speech-to-text capabilities."

“The architecture of the Achronix Speedster7t FPGA with its 2D network on chip (NoC) and ML processor (MLP) arrays gave us the building blocks required to create an ASR product that is significantly more optimised than anything available on the market today,” said Peter Baldwin, CEO of, a company known for its expertise in optimising low-latency ML inference for real-time applications. “The extremely low latency inherent in these FPGAs makes them ideal for real-time workloads.”

The accelerated ASR solution is expected to have a revolutionary impact on industries that depend on rapid and accurate speech-to-text conversion.

Its features include compatibility with major deep learning frameworks such as PyTorch plus re-trainability for multiple languages or specialties. The solution is currently being deployed with early-access customers and is now available to the general market.