Data & models
Custom keywords and commands, dataset preparation, model development and evaluation.
SONOEDGE VOICE / EMBEDDED SPEECH AI
Offline speech AI.
Built for your hardware.
Custom offline KWS and voice AI, developed around your product, vocabulary and target platform.
For consumer electronics, semiconductor platforms, IoT devices and wearables. We work across the full pipeline, from data preparation to embedded inference and on-device validation.
TARGET-HARDWARE ENGINEERING EXAMPLE
240 MHz CPU
192 KBRAM
1 MBFlash
Speech AI delivered and validated on target hardware with constrained compute and memory.
Project-specific engineering example; requirements and performance are evaluated for each target platform.
Embedded optimization
Quantization, memory planning and inference optimization for available CPU or NPU resources.
Device integration
C/C++ inference integration, hardware bring-up support and validation in the target environment.
DELIVERED ENGINEERING EXPERIENCE
Two languages.
Different hardware constraints.
MANDARIN CHINESE
KWS on RISC-V with NPU
Mandarin keyword-spotting development on a RISC-V platform with neural-processing acceleration.
- Target architecture
- RISC-V + NPU
- Engineering focus
- Speech-model development and inference integration for the target platform.
SPANISH
KWS within a small memory budget
Spanish keyword-spotting development on a 240 MHz platform with 192 KB RAM.
- Target resources
- 240 MHz CPU · 192 KB RAM
- Engineering focus
- Resource-aware model development and embedded deployment.
Full-pipeline technical solutions are available, from data preparation and model development to embedded inference, optimization and on-device validation. Each engagement is scoped to the customer’s platform and acceptance criteria.
CUSTOM KWS & VOICE COMMANDS
A voice interface shaped
around your device.
Custom wake words
Develop keyword spotting around your chosen vocabulary, acoustic conditions and false-trigger requirements.
Offline voice commands
Recognize a defined command set locally. Select vocabulary and interaction behavior around the product’s intended tasks.
Platform integration
Adapt models and inference for CPU-only, RISC-V or NPU platforms, with compute and memory constraints considered from the start.
WHO WE WORK WITH
From silicon platforms
to finished devices.
For semiconductor teams, consumer-device manufacturers and IoT product developers.
Semiconductor & platform teams
Map speech workloads to the target architecture and define an embedded reference implementation.
Consumer electronics & wearables
Build a product-specific voice interaction with an agreed vocabulary and device resource budget.
IoT & embedded equipment
Add local command recognition where the application calls for on-device operation.
FROM DATA TO DEPLOYMENT
Define the task.
Validate on the device.
Scope the voice interaction
Keywords, languages, microphone path, operating noise and acceptance criteria.
Prepare data & develop models
Agree data coverage, evaluation splits and the model approach for the target task.
Optimize embedded inference
Quantization, C/C++ integration and profiling within the device memory and compute budget.
Validate & hand over
Evaluate on target hardware and agree integration materials, test results and support scope.
Measure what matters to your product
We agree how to evaluate missed activations, false activations, latency, memory use and compute or power cost. Results depend on vocabulary, noise conditions, distance, data and target hardware.
The 240 MHz / 192 KB RAM / 1 MB Flash example is an engineering reference, not a universal requirement or performance guarantee.