Meta acquired WaveForms for its voice research; Conneau's team joined MSL

Why it matters

WaveForms' path from a $40 million seed to Meta's AI lab shows how platform companies can acquire scarce voice-model talent before a startup proves a standalone product. Conneau's post confirms the team joined MSL, while leaving the purchase price and the technology's eventual role undisclosed.

Meta's acquisition of voice-AI startup WaveForms was reported in August 2025. On September 23rd, 2026, co-founder Alexis Conneau (@alex_conneau), in a post on X, described WaveForms' team as having joined Meta Superintelligence Labs and said some of its work would appear at Meta Connect. The post brings a founder's account of the integration, more than a year after the deal first became public.

That timing matters. The acquisition is not new, and Conneau's post does not announce a purchase price or new transaction. It does put a public voice behind the hiring outcome: Conneau thanked Meta CEO Mark Zuckerberg (@finkd) for backing the vision and giving the work the scale to become "something much bigger." The wording is a founder's account of the rationale, not a disclosure of how Meta will deploy the technology.

A startup built around voice, not transcripts

Conneau arrived at WaveForms with a research record in language and machine learning and experience at Meta's FAIR group, Google AI Language and OpenAI. He studied mathematics at École Polytechnique and completed his Ph.D. at FAIR Paris in 2019, according to Stanford's NLP group. At OpenAI, he co-created the audio systems behind GPT-4o's Advanced Voice Mode, according to Axios and the startup's backer Andreessen Horowitz.

Conneau left OpenAI to pursue a different technical premise: a model that processes audio directly, rather than a pipeline that transcribes speech, generates a text response and then converts that text back to speech. In its December 2024 investment announcement, a16z described WaveForms' aim as building an end-to-end audio language model that could respond to emotional cues in a speaker's voice. Conneau called the goal the "Speech Turing Test," Axios reported at the startup's launch.

The technical distinction was also the company's business wager. WaveForms was trying to build a foundation model for voice interaction, with emotion and conversational cues as core inputs, rather than sell a narrow transcription or voice-generation feature. That approach could support multiple products, but it required substantial research before the company had a finished product to sell. When WaveForms launched, Conneau told Axios that the team was still developing its models and had five employees.

A $40 million bet, then an undisclosed sale

WaveForms raised $40 million in seed funding led by a16z. The valuation reported at the time was not consistent across outlets: Axios reported a $200 million valuation, citing Conneau, while TechCrunch later cited PitchBook data putting the round at a $160 million pre-money valuation. Those figures use different descriptions and should not be treated as interchangeable. Meta's purchase price was not disclosed when TechCrunch reported the acquisition on August 8th, 2025, citing The Information.

The interval between funding and sale was short. WaveForms launched publicly in December 2024; the acquisition was reported about eight months later. That is a quick transition from venture-backed research company to platform-owned team. It also leaves the central financial outcome opaque: the seed amount and reported valuations describe the financing, not what investors or founders received in the acquisition.

Meta gained a team led by a researcher who had already built high-profile voice technology inside another frontier AI lab. Conneau's thanks to a16z in the thread underscores the path: the fund backed WaveForms from the start, then the founders and researchers moved into Meta's organization. The investor's original announcement laid out the research thesis; the acquisition put that thesis inside a company with the resources and distribution to pursue it at a different scale.

The work now has a platform owner

The thread identifies the destination as Meta Superintelligence Labs and thanks Zuckerberg for "sharing the vision." It does not explain which WaveForms models or research projects Meta retained, how the team is organized, or what will be shown at Meta Connect. Conneau's reference to the event is the only product-timing detail in the supplied announcement; it should not be read as a launch date or proof that a WaveForms product will ship under its own name.

Meta had already bought voice-AI company PlayAI in July 2025, according to TechCrunch, placing the WaveForms acquisition in a broader push to bring audio capabilities and specialist teams into its AI organization. The two deals make voice a clear area of investment. They do not establish that the companies' products were combined or that WaveForms technology powers a particular Meta assistant or device.

The acquisition also illustrates the difficult economics of building an independent model company. WaveForms attracted $40 million before it had public operating metrics or a widely available product; its technical thesis and team were valuable enough for Meta to bring them inside. A sale can give researchers access to infrastructure and distribution that a small startup would have to build itself. It can also end the independent company's chance to prove that its research supports a standalone product and durable revenue.

For Conneau, the new post frames the outcome as an expansion of the original ambition, from a five-person startup still developing models to work inside Meta's AI lab. The next measurable test is whether that voice research appears in a Meta product, and what capabilities Meta chooses to expose. The Connect reference points to a venue, not yet to a public technical specification or a separate WaveForms product.