Tech • AI • Robotics • Game
This AI researcher wants to rid us of keyboards - Neil Zeghidour, CEO of Gradium
6/10TL;DR
Gradium is betting that near-instant voice AI will become the default interface for machines, while warning that personal AI agents remain vulnerable to serious security failures.
KEY POINTS
A startup built around voice as the next interface
Nail Zeguidour, 35, chief executive of Gradium, says speech will ultimately replace much of today’s keyboard-based interaction with machines. The company develops voice models for transcription, speech synthesis and conversational interfaces, aiming to let businesses build agents with almost any accent, style or vocal identity. Its headline claim is speed: responses in about 50 milliseconds, close to real-time human turn-taking.
From academic audio research to industrial AI
Zeguidour has spent roughly 10 years in audio AI. He began in Paris working on how babies acquire language and how machines could learn more efficiently from limited, noisy data. He later joined Facebook’s research activity in Paris for doctoral work on speech, then moved to Google, where he helped create an audio LLM architecture that could unify multiple audio tasks the way text LLMs unified translation, completion and correction.
How audio LLMs changed the field
The technical leap was to compress sound into discrete audio tokens and train a model to predict the next segment of sound, rather than the next word. That made it possible to continue a voice after only a few seconds of input, effectively enabling voice cloning, and to extend instrumental fragments into longer sequences, opening the door to modern music generation. Zeguidour argues many current products, from synthetic voices to music generators and advanced voice assistants, rest on variants of the same underlying idea.
Smaller models, lower cost, faster deployment
Unlike frontier reasoning models that can reach around 1 trillion parameters, Gradium’s voice systems operate in a range of roughly 2 billion to 5 billion parameters. That makes them far cheaper to train and run, an important advantage for real-time speech. Zeguidour says this lower infrastructure burden gives smaller teams room to compete with much larger groups.
Why large companies can be slower than startups
Zeguidour left big tech convinced that generative AI, especially voice, was too disruptive to fit easily inside companies built around older product lines. He argues AI agents multiply the output of each individual enough that large organizations can become less efficient than smaller ones. In his view, this is an “age of gold” for human-scale companies, because compact teams can move faster and adapt products more aggressively.
Personal agents are useful, but highly risky
The same technology that can triage inboxes, watch calendars and surface urgent issues can also leak sensitive data. Zeguidour described attempts to build a personal assistant with access to email, Slack and scheduling tools, only to find it exposed to prompt injection. A malicious instruction hidden in an editable calendar invite or external message could override the agent’s rules and extract confidential information.
Security hinges on barriers and limited permissions
He said the safest current approach is to keep assistants in tightly controlled silos and avoid granting autonomous sending rights. Agents may be allowed to draft replies, summarize requests and propose actions, but a human should approve external communication. The core challenge is not only model quality but the orchestration layer above the model, where permissions, memory, tool use and guardrails determine whether an assistant is helpful or dangerous.
Assistants are already moving from novelty to operations
Despite the risks, AI agents are already proving useful for prioritization and monitoring. Zeguidour says they can distinguish critical investor outreach from low-priority requests, flag expiring domains or urgent infrastructure issues, and escalate incidents overnight. In production environments, agents can alert on-call engineers in progressively stronger ways until someone responds, showing how quickly these systems are moving into operational roles.
Voice AI could reshape physical workplaces
If employees increasingly speak to AI agents throughout the day, office design may have to change again. Open-plan layouts worked when interaction was mostly silent typing, but constant voice exchange creates noise and privacy problems. Zeguidour expects more sound-isolated booths, partitions and other acoustic adaptations as spoken interaction becomes routine.
Music disruption remains less clear than expected
Even with major technical progress in generated music, Zeguidour believes the cultural shift has been slower than many feared. Live performance still commands demand, and listeners do not necessarily form the same emotional attachment to machine-generated songs as they do to human artists. AI may still infiltrate playlists, remixing and production workflows, but fully replacing musicians appears far less immediate than the underlying technology might suggest.
CONCLUSION
Gradium is pushing voice AI toward a future of seamless, always-on interaction, but the path to mass adoption depends as much on trust and security as on raw model performance. The companies that solve both may define how people work, communicate and use software in the next phase of AI.
Ask a question