Speak is an AI-powered language learning platform that focuses on improving users’ spoken fluency through conversation-based practice. Primarily offering instruction in English and Spanish, the application operates in a mobile format and was founded in 2016 by Connor Zwick and Andrew Hsu in San Francisco. The company has offices in San Francisco, Seoul, Tokyo, Taipei, and Ljubljana.
Development and Growth
Speak was first launched in the South Korean market in 2019, where it quickly became one of the most widely used English learning applications. The platform has since expanded to over 40 countries, reaching more than 10 million users globally. As of 2024, Speak raised $78 million in a Series C funding round led by Accel, bringing its total funding to $162 million and its valuation to $1 billion. Previous investors include OpenAI Startup Fund, Khosla Ventures, Y Combinator, Founders Fund, and Buckley Ventures, along with individual investors such as Sam Altman, Peter Thiel, and Jeff Weiner.
Technological Infrastructure
Speak is structured around a speaking-first methodology. The app offers lessons and exercises that encourage users to practice speaking aloud, addressing a gap in traditional language learning platforms that often lack sufficient spoken practice.
The core technology of Speak relies on streaming ASR (automatic speech recognition). The company developed a proprietary ASR model capable of recognizing speech from beginner users with diverse accents. Trained on Speak’s own user data, this model achieved a 60% reduction in word error rate (WER). It was built using NVIDIA’s NeMo open-source framework for speech AI.
To provide real-time feedback, Speak integrates Riva and Triton Inference Server architectures within a Kubernetes-based infrastructure on the Google Cloud Platform. Speech data is transmitted using WebSocket and gRPC (Google Remote Procedure Call) protocols.
API Integration
In 2024, Speak introduced a feature called Live Roleplays, utilizing OpenAI’s GPT-4o Realtime API. This allows users to engage in real-time spoken role-play scenarios with AI agents. The system evaluates not only word accuracy but also intonation, pronunciation, and prosody, providing immediate, detailed feedback.
Learning Philosophy
Speak’s learning model is based on three steps:
- Intensive listening and speaking practice in the target language
- Repetition of learned patterns with varied expressions
- Reinforcement through AI-powered real-world simulations
Personalized lesson plans, feedback mechanisms, and goal-oriented guidance support the learning process. A proficiency graph is used to track user progress and offer level-appropriate vocabulary and sentence structures.
Enterprise Applications
In addition to individual users, Speak offers a Speak for Business service for corporate clients. This includes customized lesson content and reporting tools aimed at improving employees’ English proficiency. The program is used by over 200 corporate clients and has achieved an 85% user adoption rate.
The platform has seen rapid adoption in South Korea, Japan, and Taiwan, and has expanded into Spanish-speaking and Mandarin-speaking markets as well as North America and Europe. Its user base has doubled annually, attracting continued investor interest.
Future Outlook
Speak aims to expand its speaking-based learning system to additional languages and increase the level of personalization. While currently supporting English and Spanish, the company is developing support for languages such as French. Planned future features include phoneme-level pronunciation feedback, refined fluency scoring, and speech-to-speech models.
According to co-founder Connor Zwick, Speak’s long-term goal is to build the world’s most advanced AI language tutor to help millions speak foreign languages with confidence. To that end, Speak continues to enhance its personalized learning plans, interactive content, and technical infrastructure to play a significant role in global language education.
