Speak

Speak builds AI language tutors that let learners practice speaking aloud through structured lessons and open-ended conversations. Its teaching method follows a Learn → Practice → Apply loop: tutors explain and rehearse language, then learners use it in scenarios such as ordering at a café or talking with a coworker. Speak supplies lesson goals, progression, and feedback logic; conversational models adapt the interaction to what learners say.[9][10]

On September 10, 2026, co-founder and CTO Andrew Hsu introduced Live Tutor Lessons, powered by OpenAI’s GPT-Live-1, with a limited rollout for English and Spanish learners. Interviews and live testing had shown that many learners did not naturally interrupt a tutor to ask questions. Speak therefore designed explicit moments for the tutor to check understanding and invite answers in the learner’s own words. A lesson about work, for example, could have learners describe their own jobs and responsibilities while retaining its instructional goals.[10]

Hsu reported that configuring GPT-Live-1 as a language tutor rather than a generic assistant reduced interruptions of learner thinking pauses from 27.6% to 13.6% in an internal evaluation; no sample size was supplied for that comparison. In 27 complete sessions with a simulated learner, the model delivered 476 of 477 authored lesson lines, including lessons of up to 32 steps. A separate evaluation of 101 scored learner turns found that roughly 85% of its corrections were warranted and about 80% of targeted errors were caught. It avoided distracting corrections in all 12 evaluated cases where restraint was appropriate, but still missed errors and sometimes offered stylistic feedback on valid language. These evaluations measured tutor behavior and instructional delivery, rather than human learning outcomes.[10]

The live lessons build on an earlier progression from speaking exercises to open-ended AI tutoring. Speak’s initial focus on English learners in South Korea supplied the market context for that development.[2][4][3][7]

Connor Zwick and Andrew Hsu founded Speak in 2016. Founded in San Francisco, the company launched in its inaugural market, Korea, in 2019 and concentrated on helping learners speak English. Its early content modeled everyday conversations and gave pronunciation feedback. In November 2022, Speak reported nearly 100,000 paying subscribers, just under 15 million English lessons started that year, and more than 10 million pieces of pronunciation feedback.[2][4][6]

Speak’s November 2022 Series B announcement described an AI tutor under development for open-ended conversations and real-time feedback on pronunciation, grammar, and vocabulary. The OpenAI Startup Fund investment also brought advanced access to OpenAI systems. By March 14, 2023, Speak had used GPT-4 in production for about two months to power parts of AI Tutor, including contextual feedback. The company reported more than two million lesson starts powered partly by the model.[2][3]

Speak began international expansion in 2023 after its South Korea concentration. By August that year, it reported availability in more than 20 countries, including Japan, Taiwan, Germany, France, Brazil, and Mexico. In June 2024, it reported more than 10 million learners in over 40 countries and announced English offerings for Mandarin-speaking markets and Spanish for North America and Europe. The learner count was not defined as active users or paying subscribers.[4][6]

Recognizing beginner speech required work beyond integrating general-purpose models. In a June 2024 engineering account, staff software engineer Jose Picado described replacing separate iOS and Android speech systems with unified backend inference and post-processing. Older-device constraints had limited the size of on-device models, while third-party services struggled with heavily accented English. Speak used Nvidia NeMo to fine-tune Conformer-CTC on many thousands of hours of learner speech across native languages. Conformer combines self-attention for global audio dependencies with convolution for local dependencies; the CTC variant supported fast streaming inference and alignment between audio and output.[5]

On an internal test set containing several hours of human-labeled learner speech, Speak reported more than a 60% reduction in word error rate against pretrained Conformer and a separate 45% improvement against its earlier Android on-device model. Picado also reported an average of about 1.6 seconds from the speaking prompt and start of recording to feedback on the first spoken word—a 20% speed improvement over the previous third-party recognition service.[5]

The system served the model through Nvidia Riva and Triton in GPU-backed Kubernetes pods on Google Cloud Platform. Clients streamed audio over WebSockets, and Speak servers called Riva through gRPC. Persistent gRPC connections defeated the intended distribution of work by Kubernetes’ connection-based load balancer: one of five Riva pods received no traffic. Speak introduced a custom application-layer load balancer that distributed inference traffic across all five pods.[5]

In October 2024, Speak announced Live Roleplays after testing OpenAI’s GPT-4o Realtime API. Direct speech-to-speech reduced the need for sequential transcription, text-model processing, and speech synthesis. Speak’s learning engine added proficiency tracking, level-appropriate vocabulary and sentence patterns, objectives, and hints. The announcement planned an initial limited rollout and acknowledged weaker instruction following and limitations in nuanced pronunciation coaching.[7]

Picado’s March 2026 platform account described a hybrid approach: an ASR → LLM → TTS cascade for immersive roleplays and free-form conversations, and speech-to-speech for pronunciation feedback and coaching on delivery. Cascade offered component-level control, better text-model instruction following, and lower cost. Speech-to-speech preserved tone, prosody, and pronunciation that a transcript could lose, while reducing latency. Voice agents used LiveKit Cloud for WebRTC transport and the LiveKit Agents framework, running in Speak’s regional Kubernetes clusters.[9]

Speak selected TTS providers by language pair and use case, accounting for pronunciation, code-switching between the learner’s native and target languages, voice quality, and latency. Turn-taking also needed learner-specific design: pauses while searching for words could fragment transcripts or trigger premature responses. Some features let learners tap to finish a turn, preserving the whole utterance for transcription; hands-free features used semantic turn detection. Picado described reliable automatic turn detection for language learners as an open problem.[9]

Speak announced a $27 million Series B led by OpenAI Startup Fund on November 18, 2022. A $16 million Series B-2 led by Lachy Groom followed on August 31, 2023, with reported cumulative funding of $63 million. On June 18, 2024, Speak announced that it had raised $20 million in Series B-3 financing led by Buckley Ventures at a $500 million valuation, reporting $84 million raised to date.[2][4][6]

On December 10, 2024, Zwick announced $78 million raised in Series C funding at a $1 billion financing valuation, bringing reported total funding to $162 million. Accel led the round, with participation from existing investors including OpenAI Startup Fund, Khosla Ventures, and Y Combinator; Accel partner Ben Quazzo was joining Speak’s board. Zwick also reported more than one billion sentences spoken and over 25 million personalized lessons created that year. Speak for Business, introduced since the preceding funding round, had more than 200 enterprise customers.[8]