SANTA CLARA, Calif. - TelAve -- Sensory Inc. (https://www.sensory.com/), a pioneer in on-device AI, today announced a breakthrough in embedded speech recognition with the launch of its latest Speech-to-Text engine. Optimized for TensorFlow Lite Micro (TFLM) (https://ai.google.dev/edge/litert/microcontroll...) and advanced Neural Processing Units (NPUs), including the Arm® Ethos™-U55 (https://developer.arm.com/documentation/109267/...), this engine delivers unparalleled accuracy and performance in an ultra-compact footprint.

Sensory's STT engine supports 37 languages, enabling manufacturers to deploy global products with a single, ultra-efficient architecture.

Maximum Power, Minimum Footprint
By leveraging specialized Neural Processing Units (NPUs), Sensory's STT engine eliminates the performance-draining data transfers between a CPU and an NPU. This architecture offloads the entire tensor computation graph to the hardware accelerator, which can significantly reduce power consumption and latency. By keeping the CPU idle during inference, device manufacturers can extend battery life or reserve processing cycles for complex system tasks and UI management.

More on TelAve News

• Engineering leaders from industry, academia to gather at IISE Annual Conference & Expo in Arlington, Texas
• AI-Driven Neurotechnology Expansion as FDA Path Clears and New Defense Initiative Emerges for NRx Pharmaceuticals (N A S D A Q: NRXP)
• BestDoc Launches AI Call Center for Healthcare
• Acuvance Appoints Sandeep Sabharwal to Board of Directors, Strengthening Leadership to Support Continued Platform Growth
• Grange Insurance Association to Rebrand as Granwest Insurance on July 1, 2026

The engine is available in two optimized configurations:

• 2.7 MB Domain-Specific Model: Optimized for large vocabulary "Command & Control" tasks, this model utilizes domain adaptation to maintain high accuracy in specific environments, like automotive cabins. It features a peak SRAM usage of 787.11 KiB and operates at 892.9 Million MACs per inference.
• 13 MB General-Purpose Model: A versatile model designed to handle natural language and large vocabularies without per-domain tuning. It fits within standard 2MB SRAM limits with a peak footprint of 1.68 MB, operating at 4.37 Billion MACs per inference.

Sensory's STT engine is engineered for rapid portability across a broad ecosystem. By using LiteRT (https://ai.google.dev/edge/litert/microcontrollers/overview) Micro (formerly Tensorflow Lite Micro) as the essential runtime layer, Sensory provides seamless integration for:

• Arm® Ethos™ NPU Family: Native support for Ethos-U55, U65, and U85.
• Cadence® Tensilica® HiFi DSPs: Full compatibility with the HiFi 4, HiFi 5, and HiFi iQ series.
• Edge Platforms: Optimized for Arm Cortex-M (https://ai.google.dev/edge/litert/microcontroll...) (M4, M7, M55) and popular boards like Arduino Nano 33 BLE Sense (https://ai.google.dev/edge/litert/microcontroll...), ESP32 (https://ai.google.dev/edge/litert/microcontroll...), and Sony Spresense (https://ai.google.dev/edge/litert/microcontroll...).

Privacy and Performance
"Our STT engine demonstrates that natural language interfaces can be powerful without relying on the cloud," said Todd Mozer, Chairman and CEO of Sensory (https://www.linkedin.com/company/sensory-inc-). By processing 100% of voice data on-device, Sensory helps developers ensure user privacy, lower latency, and reliability in environments with limited or no connectivity.

More on TelAve News

• Dr. Rosendo Icochea, MD Recognized for Contributions to Surgical Education and Medical Research
• Netberg SONiC 202511.n0: Powering Next-Level Performance, Security, and AI-Ready Networks
• Giftella Launches AI Gift-Finder App That Replaces Guesswork With Personalized Picks in Seconds
• Beverly.io Announces Nationwide Expansion and Poppins Payroll Partnership for Families
• New Book: The Battle for Truth and Shadows - Guardians of Light - Epic Fantasy Unveils a War Between Light and Deception

Why Embedded STT

• Privacy & Security: Voice data never leaves the device, ensuring user privacy.
• Low Latency: Instantaneous results without relying on internet connectivity.
• Lower power/lower heat: Model efficiency and NPU usage reduce power substantially.
• Cost Efficiency: Eliminates ongoing cloud processing fees and reduces data transmission costs.
• Reliability: Performance even in "comms-denied" environments or areas with poor cellular service.