AssemblyAI lands $50M to build and serve AI speech models

Companies are betting big on generative AI to gain a competitive edge. But adoption challenges remain. According to a recent survey from EY, a significant portion of businesses looking to embrace generative AI say that the field’s rapid progress — and the surge in vendors claiming to have AI expertise — is complicating their deployment prospects.

You wouldn’t know it from their spending, though. Per an IDC forecast, worldwide investments in “AI-centric” systems could reach $154 billion by the end of the year. And an MIT Tech Review poll found that 50% of firms plan to boost budgets on data infrastructure and AI by over 25% in the coming year.

The boom is benefitting startups like AssemblyAI (which TechCrunch has covered thrice before), a self-styled “applied AI” venture that researches, trains and deploys AI models for developers and product teams to integrate into their apps and services.

AssemblyAI claims that its paying customer base grew 200% from last year to 4,000 brands and that its AI platform is now handling around 25 million API calls per day. Moreover, over 200,000 developers are building on the platform, AssemblyAI says — using it to process more than 10 terabytes of data a day.

“AI models are rapidly improving and evolving,” AssemblyAI co-founder and CEO Dylan Fox told TechCrunch in an email interview. “Enterprises that leverage AssemblyAI’s API platform are able to focus on building new AI products, applications and workflows without having to focus on model development, training and keeping up with the rapid pace of model innovation. Nor do they need to worry about deploying AI models at scale themselves, which is extremely challenging to do for low cost and with high availability.”

AssemblyAI’s success has caught the attention of big-name investors, some of whom recently contributed to a new funding tranche for the startup. Accel led a $50 million round in AssemblyAI, announced today, with participation from former Salesforce co-CEO Keith Block, GitHub ex-CEO Nat Friedman and Daniel Gross, Insight Partners and Y Combinator. AssemblyAI’s total capital raised now stands at $115 million.

Fox, a machine learning engineer by trade, founded AssemblyAI in late 2017. He says he was inspired by the Amazon Echo, which Fox argues is one of the first great examples of products made possible by better AI systems for voice.

“As I started to explore building my own products with various speech AI models available at the time, I was disappointed that most companies were still offering legacy, inaccurate speech AI models through hard-to-use developer products,” Fox said. “This motivated me to start AssemblyAI — with the vision to create superhuman … AI models, available through an easy-to-use developer platform, that would unlock entirely new classes of AI applications to be built.”

Today, AssemblyAI offers AI models — specifically speech-focused models — designed to perform tasks like speech-to-text, identifying speakers, moderating content and summarizing speech through an API. Customers like Fireflies, a meeting transcription app, run content ranging from phone calls and Zoom meetings to podcasts and videos through the models, Fox says.

Now, there’s no shortage of both open and proprietary speech models out there, from rival startups such as Deepgram, Rev and Speechmatics as well as tech giants like Google Cloud, Azure and AWS. But Fox makes the case — rightly or wrongly — that AssemblyAI’s models are more “advanced,” “accurate,” “capable” and “feature rich” than the competition.

“Big cloud companies have similar product offerings … but they’re infrequently updated, less accurate, come with way fewer features and are much harder to integrate,” he continued.

AssemblyAI isn’t resting on its laurels, that said. A slice of the new funding will be put toward a “universal” speech model that the company’s training on over a petabyte of voice data, set to launch later this year,” Fox says. AssemblyAI is also expanding its headcount, aiming to grow its 115-person workforce by 50% to 75% next year.

“We’re working toward building the ‘Stripe for AI models’ — where developers and product teams will be able to easily access state-of-the-art AI through a simple API,” Fox said. “By providing these things to customers, they can focus on building more vertical applications and internal workflows that leverage our proprietary data and AssemblyAI’s ever-improving speech AI models … We have years of runway thanks to the new funding round, and are seeing an incredible amount of demand and product adoption given the mainstream push around AI.”