A novel approach to keyword detection
"The keyword detection function converts voice commands into electrical signals, which are then processed by an integrated circuit to extract frequency characteristics."
explains Franck Badets, Research Director at CEA-Leti. "This data can be analyzed by a neural network to identify spoken words. When a keyword is detected, the circuit triggers a specific action, such as waking up another system."
To minimize energy consumption, the researchers explored innovative architectures based on injection-locked oscillators.
"Injection locking is a common property of oscillators—whether biological, optical, mechanical, or electrical,"
adds Franck Badets.
From 2017 to 2020, initial research work validated the fundamental concepts and led to funding from CARNOT, IPCEI and France 2030.
“We were then able to develop a prototype, which is now protected
by five patents.”
Optimized performance with minimal power
The use of injection-locked oscillators enables efficient signal processing, explains Ali Mostafa, Integrated Circuit Design Engineer at CEA-Leti:
"With a single step, we can divide the signal into frequency bands, extract its energy and convert it into a format that is easily processed by the neural network.”
This innovative solution uses up to four times less silicon area than existing technologies and consumes up to ten times less energy, while achieving a 91% recognition rate for ten keywords.
Another key advantage of this breakthrough is that the system can function with operating voltages as low as 400 millivolts.
“Our approach is compatible with energy recovery sources and advanced semiconductor nodes such as 22nm technology, and those of the next generation, such as the 10nm technology under development at CEA-Leti”, "
highlights Ali Mostafa.
Expanding applications beyond speech recognition
Beyond keyword detection, the system's efficient neural processing opens doors for broader applications.
. "We’re proposing a recurrent neural network that requires minimal hardware resources,"
explains Emmanuel Hardy, Edge AI Architecture Engineer at CEA-Leti. "We have optimized the trade-off between accuracy and energy consumption, and trained the network using thousands of speech samples from a well-known benchmark. Finally, we precisely modeled the energy extraction circuit and were able to fine-tune the system parameters for optimal performance."
The ability to optimize and fine-tune this innovative technology makes it possible to extend applications to other fields:
“The next phase of development will focus on adapting the technology to other fields that require monitoring capacities such as health, large-scale infrastructures and predictive maintenance,”
concludes Emmanuel Hardy.
With this breakthrough, CEA-Leti continues to push the boundaries of low-power AI, offering new possibilities for autonomous, embedded intelligence across multiple industries.