At the end of June, Speak, an AI education company invested by OpenAI, completed a round of $20 million financing, raising its valuation to $500 million. This is the third round of financing completed by Speak in a year and a half. The total amount of the three rounds of financing has reached $63 million.

Not only did Speak raise a large amount of money, but the quality of its investors was also very high. Among Speak's investors are not only OpenAI but also the well-known Khosla Ventures, which was OpenAI's earliest investment institution. In addition, YC founder Paul Graham and LinkedIn CEO Jeff Weiner also participated in the investment.

The reason why so many investors are willing to bet is that Speak's revenue has seen exponential growth.

According to DianDian Data, in February this year, Speak's global monthly revenue on the App side has grown to $1.43 million, while in November last year, this figure was only $740,000. That is to say, in just three months, Speak's revenue has doubled.

If you include web revenue, Speak's revenue has already ranked third among similar products, second only to Duolingo and Babbel.

Behind Speak's explosion, AI is affecting the education industry at a speed beyond imagination. According to data released by data.ai on May 21, among the top 20 education apps in the US App Store, 5 are AI products that help students complete school assignments.

At least for now, AI has a good chance of solving the problem of large-scale and high-quality supply of educational resources. This may be why so many big names are optimistic about the AI education track.

/ 01 / Revenue tripled in 3 months, ranking among the top three in the industry

Since 2023, Speak's performance has been stable in the top 10 "education apps" in the South Korean Apple Store. According to Beluga Data, in February, Speak's average DAU on both ends reached 107,000 (average DAU for the past 30 days), an increase of 42% compared to November last year.

Compared to traffic data, Speak's performance in commercialization is even more impressive. In February this year, Speak's global monthly revenue on the App side has grown to $1.43 million, double that of November last year.

According to Statista data, if you include web revenue, Speak's revenue has already ranked third among similar products, second only to Duolingo and Babbel. Even in the Korean market where Speak focuses, its revenue has surpassed Duolingo.

From a product positioning perspective, Speak focuses on spoken language learning, allowing users to practice English conversation through a series of interactive speaking experiences. Under the guidance of AI Tutor, users can continuously practice speaking as if they were conversing with a real teacher, and receive real-time feedback on pronunciation, grammar, and vocabulary.

Before starting the teaching, Speak will ask users to make some choices, including the purpose of learning English, the skills they hope to improve, their current English proficiency, topics of interest, and the amount of time they plan to study each day. Then, the App will generate a set of courses that matches the user's level and interests based on their choices.

Speak's courses are mainly divided into video courses, speaking practice, and role-playing. In the video course, it will play a recorded video for users to watch, explaining some basic grammar and usage in detail. In the speaking practice section, users need to read aloud the example sentences, and then the AI will use voice recognition to determine if the user's reading is correct and provide real-time feedback to the user.

After completing the video course and speaking practice, it comes to the role-playing section. Role-playing is an AI Tutor integrated with GPT-4 capabilities, which is also Speak's main selling point.

The role-playing interface is like a chat room, where AI conducts simulated situational dialogue and mimics a real person talking to you. During the conversation with AI, if you don't know how to chat, it will give you some suggestions and prompts. Throughout the process, users can engage in open-ended conversations with AI on a series of fixed topics while receiving feedback on their pronunciation, grammar, and vocabulary.

Speak said in its official blog that since its launch in the Korean market in 2019, it now has over 10 million learners in more than 40 countries. In the past five years, the number of learners has doubled every year. In South Korea, nearly 6% of Koreans are using Speak to learn English.

/ 02 / Speak did two things right

Since the emergence of ChatGPT, there have been many AI education products released. Why did Speak ultimately stand out? This is due to two things Speak did right: embracing large models earlier and investing more resources in curriculum design.

First, Speak was almost one of the earliest companies in the education industry to embrace large models. After receiving investment from the OpenAI Startup Fund in 2022, Speak integrated OpenAI's new automatic speech recognition model, Whisper, in March last year. Whisper has near-human robustness and accuracy in English speech recognition, regardless of the user's accent (such as Korean or Japanese users), making it more accurate in recognition and providing more accurate feedback.

Then, they also became one of the first language learning platforms to use GPT-4. Compared to previous models, GPT-4 has two upgrades:

First, GPT-4 can understand more complex expressions and generate more accurate and coherent text. Second, GPT-4 has a better memory and won't forget previous conversations after a few rounds, making the interactive experience smoother and making human-computer conversations highly personalized.

Embracing OpenAI earlier gave Speak's AI technology an advantage over most competitors.

In the opinion of Speak CEO Zwick, highly personalized and contextual feedback can only be developed using GPT-4. He said, "At the bottom level, we combine OpenAI's latest technology with internal models to provide optimal performance in speech recognition, speech generation, and dialogue generation, which in turn improves user engagement and learning outcomes."

In addition to having no shortcomings in AI technology, compared to many startup AI education companies, Speak has also invested more resources in curriculum design.

Before 2023, Speak focused its business on curriculum development and has accumulated nearly 15 million English courses. In teaching and research design, Speak has its own unified principles, such as making users speak within the first 30 seconds and considering the synchronized learning effect of vocabulary, grammar, and speaking in curriculum design.

In terms of curriculum design, Speak emphasizes localization. Each region of Speak has a dedicated teaching and research leader. Early content was based on the Korean curriculum version, and subsequent courses will consider both the general version and the refined version, always balancing market expansion and teaching effect improvement. In this process, Speak will continuously conduct A/B testing to iterate and correct the curriculum logic, unlike many English apps that focus most of their attention on marketing.

Through more targeted curriculum design, Speak has formed a clear contrast in experience and effectiveness with free AI speaking coaches on the market.

Since the launch of ChatGPT, there have been numerous AI speaking coach products, such as the AI language teacher virtual character on C.ai. However, due to the lack of reasonable prompts and professional training data, most AI speaking coaches perform poorly in handling professional terms, dialects, and slang, and have a higher chance of providing incorrect information, only meeting initial needs.

Overall, Speak's success is a typical case of combining vertical scenario accumulation with large model technology. Behind this, it also reveals the essence of AI education: achieving low-cost large-scale supply of high-quality educational resources through AI technology.

/ 03 / Education is being dominated by AI

In the past year, you can clearly feel that compared to other industries, AI's penetration into the education industry has been more rapid.

According to data released by data.ai on May 21, among the top 20 education apps in the US App Store, 5 are AI products that help students complete school assignments.

According to incomplete statistics, in the second half of 2023 alone, more than ten online language learning startups completed financing, with a total financing amount of $160 million. Among the investors are well-known capital such as OpenAI and Gradient Ventures, a venture capital firm under Google.

The reason why AI runs so smoothly in the education industry is that the characteristics of education highly match the current capabilities of AI.

For a long time, the education industry has been addressing one problem - how to achieve large-scale and high-quality supply? From online recorded courses to educational tool products and then to Vipkid's model of supplying "North American teachers" and online teaching, all aim to achieve this goal.

For learners, the best way to teach is definitely one-on-one service, but the cost is often the highest. With the maturity of AI technologies such as large models, AI can simulate real-time interactive teaching that was only provided by real teachers in the past. That is to say, through AI teachers, the education industry has used technology to achieve large-scale teacher supply, making low-cost real-person teaching possible.

In addition to a clear path, the barriers to implementing AI in educational scenarios are lower. First, the teaching content material itself has undergone repeated refinement and is already quite mature, with plenty of data for AI to learn from. Second, users' needs in educational scenarios are also relatively clear, which is more conducive to the design and implementation of AI products.

It must be emphasized that the current capabilities of AI in educational scenarios have not been fully developed. With subsequent iterations of large models, there is still much room for improvement in the experience and capabilities of AI education. For example, GPT-4o, the latest version of OpenAI, has already proven that AI assistants can engage in voice interactions with emotions, enabling more personalized teaching.

Perhaps there is still debate about what role AI should play in education. But one thing is certain, as OpenAI founder Altman said, AI will ultimately reshape the education industry.