AssemblyAI is a team of researchers, engineers, and designers that have worked at some of the largest technology companies all over the world. Funded by Y-Combinator, the company’s core offering is speech recognition technology offering higher accuracy by building custom speech recognition models tailored to the customer’s specific use case. This unique solution enables customers to transcribe dozens of audio files in parallel in batch mode, or transcribe audio streams in real-time for interactive applications.
AssemblyAI offers its customers an on-premises engine that customers can license and run on their own servers for privacy and latency sensitive applications. Additionally, they offer a secure cloud API for fast transcription.
AssemblyAI needed their team to have access to reliable, cost-effective hardware for the development of their Speech-to-Text application. The company was faced with the common dilemma of purchasing their own hardware and running it on-premises, hosting it off-site, or utilizing costly virtualized cloud services through a third-party big tech cloud provider.
Dylan Fox, chief executive officer for AssemblyAI states, “We’ve tried running our own GPU servers on-premises, but servers kept going offline due to a myriad of problems like disk issues, power surges, or broken GPU cards. This meant our team had to spend time debugging hardware versus focusing on our core competencies. Using other cloud services like AWS or GCP was very costly and the support was not great.”
AssemblyAI needed to ensure whatever solution they chose would give their team the ability to focus on their core competencies of model building and training. Mr. Fox continued, “We needed to ensure our team wasn’t spending time maintaining and debugging GPU infrastructure so they could stay focused and move faster, enabling us to improve the accuracy of our models faster.”
Cirrascale Cloud Services® worked closely with AssemblyAI in creating a dedicated, multi-GPUolution that fulfilled all of the requirements set forth by the customer. “For large scale, continuous training we found the big tech clouds to be expensive and inflexible,” stated Mr. Fox. “However, with Cirrascale we have a variety of hardware configurations and GPU cards we can configure to best suit our needs and budget. Plus, the level of support and customer experience is much better working with Cirrascale – it feels more like a complete managed solution.”
Ultimately, AssemblyAI selected Cirrascale’s cloud platform due to Cirrascale’s ability to provide a dedicated multi-GPU platform that could scale and meet AssemblyAI’s needs. “The hardware options are great and very flexible, the cost is very affordable, but a main reason was the support and
customer experience we received from Cirrascale,” states Mr. Fox. “They really cared about our account, which made us feel like we were working with a true partner who would help us solve our problems as we grow.”
AssemblyAI also utilized Cirrascale’s ability to provide various configurations to easily allow for benchmarking varying models. “Our customers span across the following major industries: contact centers, broadcast media, entertainment, medical, and legal. We provide speech recognition technology to companies in these industries in order to power the features within their products and services,” states Mr. Fox. “Our primary goal is to improve the accuracy of our speech recognition models. This requires a lot of model training, experimentation, and GPUs in order to iterate fast enough. With Cirrascale GPU servers, we’re able to experiment and iterate quickly with a large number of GPUs at an affordable cost and without having to manage a fleet of GPU servers ourselves.”
Download the AssemblyAI Case Study to learn more about Cirrascale Cloud Services and its cloud-based solutions.