This podcast features Andrew Su, co-founder and CTO of Speak, discussing the company's journey and vision for AI-powered language learning. Su, a former Teal fellow, shares how Speak navigated early challenges by focusing on a compelling future vision where AI revolutionizes language acquisition. Initially gaining traction in South Korea with a premium, custom-content English learning app, Speak significantly expanded its capabilities and reach with the advent of large language models (LLMs) like OpenAI's Whisper and GPT-3.5 Turbo. The conversation delves into Speak's unique pedagogical approach centered on speaking practice and functional fluency, its strategic market entry, the development of AI-generated curriculum and a "fluency knowledge graph," and the ongoing development of real-time voice capabilities. Su emphasizes the importance of practical application, human connection, and leveraging AI for massive scalability in content creation and personalized learning.
Key Takeaways from the Interview with Andrew Su (Speak CTO):
0:15 Early Career & Teal Fellowship: Andrew Su was part of the inaugural 2011 Teal Fellowship class, a controversial but life-changing experience that allowed him to drop out of grad school and pursue startups, where he met his current co-founder.
2:45 Genesis of Speak & Vision: Founded in 2016, Speak's core vision was driven by the belief that speech and language models would become superhuman within 5-10 years, fundamentally changing how people learn languages through pure software AI tutors.
3:48 Painful Early Years & No Pivot: The initial years were challenging, taking much longer than expected to build a product and find product-market fit (PMF). Despite difficulties, Speak never pivoted from its core vision.
11:00 AI Native Product Category: Speak positions itself as Gen 3 of language learning, moving beyond traditional apps like Duolingo (Gen 2, mobile/gamified) by focusing on functional fluency, real-world role plays, and spontaneous speaking, rather than just vocabulary and grammar drills.
11:30 Pivotal Re-launch in 2018: The company completely rebuilt its app, abandoning a free, undirected experience. They focused specifically on teaching English in South Korea, introduced structured courses, new lesson types, and adopted a premium-only model.
13:18 Impact of LLMs & Whisper (Act 2): The release of Whisper and LLMs like GPT-3.5 Turbo in 2022 marked "Act 2" for Speak, enabling a shift from a supplemental "listen and repeat" tool to a full-featured language tutor offering direct, semantic feedback.
13:55 Scale and Market Reach: Speak is the largest English learning app in South Korea, with 6% of the Korean population having tried it. It's also growing in other Asian markets (Japan, Taiwan) and expanding globally, with over $50 million in Annual Recurring Revenue (ARR).
14:46 B2B Expansion: Speak launched a B2B segment a year ago, which is growing rapidly and becoming a meaningful part of the business, extending beyond language to management and communication skills.
15:15 Translation vs. Learning: Su argues that real-time translation (like the "babblefish") won't replace language learning because people learn languages for personal growth and human connection, not just literal translation. Speak plans to integrate translation features into its learning experience.
16:55 Strategic Focus on Korea: Choosing South Korea was serendipitous but proved strategic due to the population's obsession with learning English and the abundance of human-based solutions, allowing Speak to test its strong PMF against established competitors.
19:10 Attention to Localization: Speak invested heavily in precise localization, including button text and language strings, leading users to believe it was a Korean company, which was crucial for adoption.
19:21 Custom ASR & In-house Content: Speak developed custom Automatic Speech Recognition (ASR) models for non-native English speakers before LLMs and maintained fully in-house content development based on their "Speak method" of teaching sentence patterns.
29:10 AI Content Generation & Knowledge Graph: Speak is actively building AI-powered agents (tutor, curriculum writer) and an LLM-based pipeline to generate content at 100x scale and support 100x more language pairs. They are also developing a "knowledge graph" to quantify user fluency (the "Speak Score"), tracking vocabulary, patterns, and mistakes.
35:37 Focus on Casual Conversational Language: Speak's fundamental tenet is to teach "real people" conversational language and slang, not textbook English, adjusting for regional dialects like American English.
37:47 Prioritizing Speaking & Overcoming Psychological Barriers: The core product encourages users to speak freely and make mistakes in a private, non-judgmental AI environment, which helps overcome the psychological barrier of speaking in front of humans.
38:55 Pronunciation Coach: Speak offers a pronunciation coach for English (fine-tuned Meta Wave2 model) for single words, with plans to expand to sentences and more languages.
46:50 Real-time Voice Platform (Future): Speak is building a real-time voice platform using OpenAI's API, aiming for multilingual capabilities with code-switching, though cost remains a key consideration for mass production.
51:00 Latency Challenges: Su highlights that perceived latency in voice interactions is dominated by voice activity detection (VAD) and turn detection, especially for language learners who hesitate, requiring custom solutions beyond standard models.
52:10 AI in Engineering: Speak encourages its engineering team to use AI coding tools (e.g., Cursor, Warp) as a default path, actively hiring for individuals with "higher agency" in trying these tools.
53:23 Future Vision Beyond Language: Speak's longer-term ambition is for AI to reinvent how people learn anything, viewing language learning as a beachhead for a universal AI tutor.
56:23 speak.com Acquisition: The company acquired the speak.com domain several years ago for a significant sum, seeing it as crucial for building a strong, globally scalable consumer brand.
1:02:10 AI's Real-World Inertia: Su believes that despite AI being the most transformative technology, its material impact on most people's lives outside tech hubs remains close to zero due to "real-world inertia," emphasizing the need for more builders of AI-native applications.