Speakeasy Labs Inc., a startup with an artificial intelligence app that helps users learn English, today announced that it has raised $78 million in funding.
Accel led the Series C investment. It was joined by OpenAI Startup Fund, Khosla Ventures, Y Combinator and other returning backers. Speak is now valued at $1 billion, double what it was worth following a $20 raise million in June.
"So far this year, users have already spoken more than one billion sentences with Speak," Speak co-founder and Chief Executive Officer Connor Zwick detailed in a blog post. "And in 2025, we plan to bring Speak to many, many more countries."
Speak's namesake mobile app enables users to learn English by engaging in conversations with an AI voice. The software analyzes the user's responses and provides linguistic feedback. According to Speak, its approach enables consumers to practice more than 1,100 phrases within a week of signing up.
The company offers an enterprise edition of its app, Speak for Business, that provides additional features. There are language courses geared towards users with varying levels of proficiency. The app also makes it possible to practice specific business conversations, such as discussions with suppliers and customers.
One of the latest additions to Speak's feature set, Live Roleplays, rolled out a few weeks ago. It automatically adjusts the AI voice's sentence patterns and vocabulary based on the user's proficiency level. Additionally, the feature displays study goals and language hints to keep the learning process flowing smoothly.
Under the hood, Live Roleplays is powered by OpenAI's Realtime API. Introduced earlier this year, the service enables applications to process voice input from users and generate responses using GPT-4o. It promises to reduce response latency by skipping several of the steps usually involved in processing audio.
Shortly before releasing Live Roleplays, Speak overhauled its app's speech recognition engine. It upgraded the engine to Conformer-CTC, a speech recognition model introduced by Google LLC researchers in 2020.
Conformer-CTC combines the Transformer architecture, which underpins most large language models, with a convolutional neural network. That's a type of AI typically used for computer vision tasks. According to Conformer-CTC's developers, the model can process speech more accurately than algorithms based solely on the Transformer architecture.
When it rolled out Conformer-CTC to its speech recognition engine, Speak detailed plans to develop custom LLMs down the line. Building such models can incur significant costs. The $78 million funding round announced today could make it easier for the company to balance those expenses with growth investments.
Elaborating on its development roadmap, Speak detailed in today's funding announcement that plans to add support for more languages beyond English. The company will start with Spanish and French next year. Additionally, Speak is working on features that will enable its app to test users' language proficiency in a more accurate manner.