Tech • AI • Robotics • Game
This AI cuts you off when you speak (and it's deliberate)
7/10TL;DR
A Paris team behind Moshi has helped redefine voice AI with a real-time full-duplex architecture that many major companies now use as a foundation for next-generation assistants.
KEY POINTS
From finance to speech research
Nail, now co-founder of Gradium, began in mathematics for finance before switching to machine learning and joining the MVA AI program in France. His early research focused on how babies learn language with limited exposure, a contrast with speech systems that once needed 30,000 hours of labeled audio and now use tens of millions. That work was carried out between ENS and Facebook AI Paris, where the goal was to understand why humans learn speech so efficiently from sparse, unlabeled input.
Turning audio into tokens
Later work on neural audio compression led to a major shift. Traditional formats such as MP3 and Opus rely on hand-designed psychoacoustic rules, but newer models learned a far more efficient representation directly from data. The result was an audio tokenizer that compressed sound into code-like units, making speech and music predictable in sequence form, much as text models predict the next word.
A key step toward generative audio
Once audio was tokenized, a language-model approach could be applied directly to sound. Feeding a few seconds of speech allowed the system to continue in the same voice, effectively enabling early voice cloning; feeding piano notes let it continue a melody. That work also opened the way for text-to-audio generation, speech synthesis, translation and sound-effect creation, and it indirectly influenced companies such as Suno, whose early open-source stack drew on the same paradigm.
Why voice assistants have lagged
Despite headline demos such as Google Duplex in 2018, voice assistants remained limited because most systems were built as cascades: speech recognition, then a text model, then speech synthesis. That approach adds latency, strips away tone and hesitation, and forces turn-taking like a walkie-talkie. Users must adapt to the machine, speaking at the right pace and avoiding interruptions, rather than conversing naturally.
The Moshi breakthrough
The answer developed in Paris was full duplex voice interaction, where the model listens and generates continuously, including silence, instead of switching between listening and speaking. Released in 2024, Moshi was built to handle interruptions, overlaps, hesitations and emotional cues more like a human conversation. A live public demo showed the system operating in real time on a laptop, a technical feat that drew strong reactions from specialists who understood the difficulty of training the tokenizer, text model and duplex interaction together.
Open source influence beyond the product
While Moshi itself was not initially a full consumer assistant, its technical impact spread quickly. The team published a detailed paper and open-sourced models, code and the tokenizer, which has been downloaded at a rate of about 1.5 million times per month. That stack has reportedly influenced systems at Meta, Mistral, Google and others, including speech architectures behind later commercial products.
Gradium’s commercial push
After seeing outside companies build valuable products on top of the open models, Nail and colleagues launched Gradium to turn research prototypes into production-grade voice systems. The company’s stated goal is to become the world leader in speech models across tasks. It has raised €60 million, arguing that a small Paris team can still outperform larger rivals in voice if it stays ahead on both architecture and product execution.
Small models, fast responses
A core part of the strategy is efficiency. Voice products need low latency and low cost, which makes giant reasoning models impractical for routine tasks such as booking appointments. Gradium is pursuing hybrid systems in which a small, fast voice model handles natural conversation and calls a larger text model only when deeper reasoning or web retrieval is required, then inserts the result smoothly back into the dialogue.
What comes next for voice AI
The company expects voice use to expand sharply by 2026, not only for personal assistants but also for robotics. Humanoid robots in factories, streets or homes will need speech interfaces because they cannot rely on keyboards, yet current models still struggle in noisy environments with multiple speakers and overlapping sounds. That problem remains unsolved, and researchers argue the final stretch is harder because audio quality and conversational naturalness still require heavy human evaluation rather than simple automated benchmarks.
Europe’s strategic opening
The broader claim is that Europe still has a real opening in voice AI, even as the US and China dominate much of the text-model race. Paris has become a major hub, with research teams from Meta, Google, Anthropic, OpenAI and Mistral operating there. Supporters of the sector argue that sovereign capabilities in end-to-end AI, including speech, are becoming strategically important for industry, health, defense and software infrastructure.
CONCLUSION
Voice AI is moving from novelty toward infrastructure, and the contest is no longer just about sounding human but about building fast, useful and economically viable systems. If the current trajectory holds, the next wave of assistants may be shaped less by flashy demos than by the quiet architecture work now spreading from Paris across the industry.
Ask a question