Skip to content

VoiceHub

Sensory's VoiceHub is a web portal that provides a convenient interface for developers to prototype and experiment with wake words, language models and natural language understanding. Users can build custom wake words, voice control command sets, and create grammar-based language models with flexible intents and entities.

VoiceHub uses Sensory's TrulyHandsfree for wake words and spotted commands, and TrulyNatural for grammar-based recognition with natural language markup to identify intents and entities.

Output format selection

VoiceHub can deliver recognizer models in various formats, as specified by the Output Format selector. If you want to use such a model with the TrulyHandsfree or TrulyNatural SDKs you should select the THF/TNL SDK: snsr file option. This is the default for new projects.

If you are using TrulyHandsfree or TrulyNatural on a small embedded platform, you should select one of the alternate output formats described below.

THF/TNL SDK: snsr file recommended
This is the standard TrulyNatural model format. Use this unless you will be running the model on an embedded platform with limited CPU cycles and available RAM.
THF/TNL SDK: snsr file (low memory use)
This optimizes the model for small platforms with limited CPU cycles (< 500 MHz) and RAM (< 1 MiB of heap). The reduced heap and CPU requirements come at the expense of a bit of recognition accuracy.
THF/TNL SDK: .c file (low memory use)
Similar to THF/TNL SDK: snsr file (low memory use) above, but also includes a model converted to C code that you can compile into your application. On platforms with read-only / flash memory this reduces the amount of RAM required by the size of the model file.
VoiceHub uses snsr-edit to create two C files from the snsr model:
  • snsr-edit -c voicehub -t model.snsr to create model.c, and
  • snsr-edit -i -t model.snsr -o model-custom-init.c to create model-custom-init.c. This file includes custom initialization code that elides unused modules at link time to reduce overall application size.
THF/TNL SDK: .c file (low memory use for ST Micro STM32H7)
Similar to THF/TNL SDK: .c file (low memory use), but also includes TrulyNatural SDK libraries for use on the STMicroelectronics STM32H7 series microcontrollers.
Embedded: Arm Cortex-M55/M85 Ethos-U55-128
Similar to THF/TNL SDK: .c file (low memory use), but with inference optimized for Arm's Cortex-M55/M85 and Ethos-U55 NPU with Vela accelerator config ethos-u55-128.
Use this on the Alif Ensemble family of microcontrollers.
Embedded: Arm Cortex-M55/M85 Ethos-U55-256
Similar to THF/TNL SDK: .c file (low memory use), but with inference optimized for Arm's Cortex-M55/M85 and Ethos-U55 NPU with Vela accelerator config ethos-u55-256.
Use this on the Alif Ensemble family of microcontrollers.
Embedded: Infineon Arm Cortex-M55/M85 Ethos-U55-128 (model in RAM)
Similar to THF/TNL SDK: .c file (low memory use), but with inference optimized for Arm's Cortex-M55/M85 and Ethos-U55 NPU with Vela accelerator config ethos-u55-128.
The compiled model code runs from RAM. This requires more available heap than the (model in ROM/Flash) option below. Use this only if the flash read speed is too low to allow the model to run in real time.
Use this only on Infineon microcontrollers.
Embedded: Infineon Arm Cortex-M55/M85 Ethos-U55-128 (model in ROM/Flash)
Similar to THF/TNL SDK: .c file (low memory use), but with inference optimized for Arm's Cortex-M55/M85 and Ethos-U55 NPU with Vela accelerator config ethos-u55-128.
The compiled model code runs from code space.
Use this only on Infineon microcontrollers.