The speech intelligence layer for developers
HQ: San Francisco, CA, United States | Founded: 2017 | Employees: About 72 (founder-stated; unusually lean for its scale) | Stage: Series C and beyond — about $160M+ raised; last announced round was a $50M Series C in December 2023 | Website: https://www.assemblyai.com
AssemblyAI builds speech AI models and exposes them as developer APIs. Its core products turn audio into text with high accuracy, then layer on what it calls Audio Intelligence: speaker labels, chapters, sentiment, topic detection, summarisation and content moderation. Developers use these to add transcriptions, meeting notes, call analytics and brand-safety checks to their products without building speech models themselves.
The company was founded in 2017 by Dylan Fox and went through Y Combinator's first AI-focused batch, at a time when most investors were still sceptical that AI could be a business. It raised a Series A led by Accel, a Series B, and a $50M Series C also led by Accel in December 2023, bringing total funding to $115M. It has kept its team small and its focus narrow: speech, and the intelligence built on top of it.
AssemblyAI's differentiator is accuracy on real-world audio — noisy phone calls, multiple speakers, accented speech, background noise — plus a broader feature set than a raw transcription service. Its Universal model is trained on more than 10 million hours of audio, and its LeMUR product lets developers run large language models directly over recognised speech to answer questions or generate structured output.
Customers include meeting-notes and media companies, contact centres and workflow tools. AssemblyAI increasingly positions itself as the speech intelligence layer that voice agents, meeting assistants and media platforms build on, rather than competing to build the consumer products themselves.
AssemblyAI has quietly become one of the most useful pieces of plumbing in the voice AI boom. While the market focuses on flashy voice assistants, AssemblyAI does the less glamorous work that everything else depends on: converting speech into accurate text, then turning that text into summaries, insights and searchable data. It processes about two million hours of audio a day, and its customers are the apps millions of people use without ever knowing the transcription layer behind them.
The company's discipline is what stands out. Founded in 2017 by Dylan Fox out of Y Combinator's first AI batch, it raised steadily — a $50M Series C led by Accel in late 2023 brought total funding to $115M — and then largely stopped needing to. By 2026 the founder was reporting $160M+ raised, revenue tripled, roughly 72 employees and close to two million hours of audio processed daily. That is a very high revenue-per-employee ratio for an infrastructure company.
The strategic question is where speech sits as the models get better. The big foundation labs offer transcription, and voice-generation companies like ElevenLabs are moving into the same customer conversations. AssemblyAI counters with a focused stack: models tuned for accuracy in messy, real-world audio, plus higher-level features such as summarisation, content moderation and an LLM layer (LeMUR) that answers questions over audio. It is the difference between raw transcription and a speech intelligence layer.
Voice is becoming a default interface — contact centres, meeting notes, media, accessibility — and that gives AssemblyAI a long runway. The risk is that the value of raw speech recognition keeps falling while the value moves to voice generation and agents, areas where larger and better-funded players are competing hard. AssemblyAI's bet is that accuracy, price and developer trust at the infrastructure layer are their own durable business.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki