WaveForms aims to pass the speech Turing Test

If you have ever tried talking to a standard voice assistant, you know the frustration. You can sound sarcastic, exhausted, or furious, and it responds with the same cheerily flat tone. That is because most voice bots do not actually listen to you; they turn your speech into text, strip out your tone, and read back a script.

WaveForms, a new startup founded by Conneau—a key contributor behind ChatGPT's Advanced Voice Mode—wants to end that mechanical stiffness. The company officially launched with an ambitious benchmark: passing what it calls the Speech Turing Test.

Moving past speech-to-text

The fundamental flaw with legacy voice assistants is the reliance on intermediate speech-to-text conversion. The second spoken words are flattened into plain text, critical human signals like sarcasm, hesitation, pitch, and emotional weight completely disappear.

WaveForms is designing AI models that skip the transcription bottleneck to grasp the emotional context directly from audio. Conneau's goal is to produce conversational agents that can interpret emotional nuances in real time and respond with humanlike cadence, making voice interaction practically indistinguishable from talking to an actual person.

The consumer-first roadmap

Instead of starting as an enterprise backend vendor, WaveForms plans to roll out consumer-facing applications first.

The logic is straightforward: nuanced, empathetic speech models need diverse, messy, real-world conversational data to get better. By deploying consumer products, the team can refine tone and emotional feedback loops before packaging the tech for enterprise use.

Down the line, the company plans to take the refined platform into education, healthcare, and customer service, where empathetic communication actually moves the needle.

Why it matters

For everyday users, while there is no standalone app to download today, WaveForms signals that voice interaction is finally moving away from robotic readouts toward genuine audio presence. If you care about AI companions or realistic voice interfaces, this is the architecture to watch.

For the broader industry, WaveForms proves that the proprietary know-how behind frontier voice systems like OpenAI's Advanced Voice Mode is already dispersing across the ecosystem. That will put pressure on incumbents like OpenAI, Google, and audio specialists like ElevenLabs to speed up their own real-time emotional audio engines.

We think launching with consumer products is a sharp move, because realistic empathy cannot be simulated in benchmark suites; it has to survive real human conversation. But the trade-off is massive. When an AI voice sounds entirely human and responds to emotional vulnerabilities convincingly, issues around emotional dependency, impersonation, and social manipulation get ten times harder to manage.