How does Speak differ from traditional language learning apps like Duolingo or Babbel?

Speak focuses on 'Gen 3' AI-native language learning, emphasizing functional fluency through speaking practice, role-plays, and sentence pattern drills rather than just vocabulary and grammar. It aims for spontaneous and automatic speech.

What was the crucial technological advancement that elevated Speak's capabilities?

The release of advanced AI models like OpenAI's Whisper for speech recognition and LLMs like GPT-3.5 Turbo in 2022 were pivotal, allowing Speak to transition from a supplemental practice tool to a full-featured AI language tutor.

How does Speak handle onboarding new users, especially those who might not be able to speak aloud immediately?

Speak uses an LLM-based conversational onboarding called 'Magic Onboarding' to personalize the experience. They also experiment with fallback options for users in situations where speaking is not feasible.

What is Speak's strategy for content generation and scaling to new languages?

Speak is heavily investing in AI content generation through agents for curriculum writing and lesson creation. This approach aims to produce 100x more content and support expansion to many more languages.

How does Speak measure and improve user fluency?

Speak focuses on 'real-world fluency' and aims to quantify it with a 'Speak Score'. This involves building a knowledge graph of a user's vocabulary, grammar patterns, and mistake history to provide a holistic view of their proficiency.

What challenges did Speak face in its early development before the advent of LLMs?

In the early years (2017-2019), Speak struggled to find product-market fit and build a great product. They developed custom speech recognition models and focused on user data to fine-tune them, persevering through painful iterations.

How does Speak address the use of casual vs. formal language in its curriculum?

Speak intentionally avoids teaching textbook language. They focus on teaching colloquial, casual conversational language that real people actually use, which is often very different from traditional textbook examples.

What are the challenges with multilingual real-time voice communication?

Handling multilingual conversations and code-switching (mixing languages within a sentence) is difficult for TTS models. Ensuring natural pronunciation and smooth transitions between languages, especially for learners, is a frontier problem.

What is Speak's stance on pronunciation and accent training?

While emphasizing that communication is orthogonal to perfect pronunciation, Speak offers a pronunciation coach for English. However, their primary focus remains on increasing speaking volume and confidence in communication.

What is the future long-term vision for Speak beyond language learning?

Speak believes AI will fundamentally transform how people learn *anything*. Their long-term ambition is to apply their AI-driven learning platform to various academic subjects and skills, not just languages.

Why is real-world inertia significant in the context of AI's impact?

Despite transformative AI technologies like GPT-4, the day-to-day lives of most people outside tech hubs haven't changed dramatically. Real-world inertia is enormous, highlighting the need for more builders creating practical applications.

Key Moments

Personalized AI Language Education — with Andrew Hsu, Speak

Latent Space Podcast

Science & Technology4 min read65 min video

Jul 11, 2025|1,953 views|52|4

Save to Pod

Key Moments

TL;DR

Speak is revolutionizing language learning with AI tutors, focusing on fluency and conversational skills.

Key Insights

Speak aims to create AI language tutors that are more effective than human teachers by leveraging advancements in speech and language models.

The company prioritizes functional fluency and spontaneous speaking over traditional vocabulary and grammar drills.

Speak's 'magic onboarding' uses AI to create a personalized conversational experience for new users.

The platform has evolved from a supplemental practice tool to a full-featured AI tutoring experience, significantly enhanced by LLMs and models like Whisper.

Speak is expanding beyond English and South Korea to offer a wider range of languages and enter new markets, including B2B offerings.

The long-term vision extends beyond language learning to reinventing how people learn anything using AI.

FOUNDING VISION AND EARLY STRUGGLES

Speak was founded in 2016 with a bold vision: to leverage the predicted superhuman capabilities of future speech and language models to create AI-powered language tutors. The co-founders, Andrew Hsu and his partner, spent a year deeply researching AI, convinced that software alone could revolutionize language acquisition. Despite this strong conviction, the initial years were challenging, marked by a lack of product-market fit and a longer-than-expected development cycle. This foundational belief in the future of AI's potential in education, however, kept the team motivated through the difficult phases.

EVOLUTION OF THE PRODUCT AND TECHNOLOGY

Initially, Speak focused on custom speech recognition models, especially before the widespread availability of models like Whisper and LLMs. The platform collected vast amounts of non-native English speaker data to fine-tune its models, focusing on a fast, low-latency ASR loop crucial for speaking practice. The integration of LLMs significantly enhanced the platform, enabling more open-ended tutoring, semantic feedback, and a more conversational onboarding experience, dubbed 'magic onboarding'.

FOCUS ON FUNCTIONAL FLUENCY AND USER EXPERIENCE

Speak distinguishes itself by prioritizing functional fluency and spontaneous speaking over rote memorization of vocabulary and grammar. The app emphasizes role-playing scenarios, mimicking real-world interactions like ordering from an Uber driver. This approach is designed to make speaking automatic and natural, akin to practice in a gym. The company also critically re-evaluated its product strategy, moving from a free, unfocused experience to a premium model and a more guided, 'on-rails' learning path to reduce user decision fatigue.

MARKET TRAJECTORY AND GLOBAL EXPANSION

Speak achieved significant success in South Korea, becoming the largest English learning app there. This initial focus on a specific market allowed the company to refine its product and gather crucial data before expanding. Beyond South Korea, Speak is actively growing in other Asian markets like Japan and Taiwan, and is launching in the US and other countries with multiple languages. The company has also seen considerable success with its B2B offerings, which started as an experiment but are now a significant growth area.

THE ROLE OF ADVANCED AI AND FUTURE AMBITIONS

The advent of models like Whisper and advancements in LLMs have been transformative for Speak, enabling a shift from simple listen-and-repeat exercises to sophisticated AI tutoring with direct feedback and explanations. The company is actively exploring further AI integration, including AI-driven content generation for scaling lesson creation across more languages and leveraging knowledge graphs to quantify fluency. The long-term vision is to apply these AI learning principles beyond languages to virtually any subject, fundamentally changing how people learn.

NAVIGATING TECHNICAL CHALLENGES AND MARKET OPPORTUNITIES

Speak faces ongoing challenges, such as controlling inference costs for real-time features and ensuring low latency, especially for multilingual interactions requiring code-switching. The company is building custom infrastructure for real-time speech and features. They also noted that while real-time translation is a related field, the user desire for language acquisition is driven by a deeper need for connection and self-improvement, rather than just translation. The B2B side extends this into enterprise learning for communication and management skills.

BUILDING TEAMS AND CULTURAL ADAPTATION

Speak's early team-building was unconventional. Initially, the company hired engineers remotely from locations like Slovenia, reflecting a global approach to talent acquisition. While this fostered a strong engineering core, Andrew Hsu acknowledges the difficulties of managing a distributed team across time zones. They also placed significant emphasis on localization and ensuring the app felt authentic to its users, leading to misconceptions that it was a local company, a testament to their attention to detail.

THE FUTURE OF LEARNING AND AI'S ROLE

The company believes AI will fundamentally reinvent learning across all domains, not just languages. They see AI as a powerful tool for personal curiosity and structured learning. The development of AI agents, curriculum writing tools, and sophisticated knowledge graphs are key to scaling their offerings. Speak's ambition is to create a 'speak score' that holistically measures fluency, enabling users to achieve real-world proficiency and meet their learning goals, extending from academic subjects to professional skills.

Mentioned in This Episode

●Products

●Software & Apps

●Tools

●Companies

●Organizations

●Books

●Concepts

●People Referenced

Common Questions

Speak's vision is to create an AI language tutor that helps users become fluent faster than humanly possible, leveraging advanced speech and language models to provide a personalized and effective learning experience.