Skip to content

Overview

This section provides a brief overview of this SDK: Features supported by variant, development host requirements, supported target platforms, snsr model files, command-line tools, and the software license keys we use to control library features.

Variants

The TrulyHandsfree, TrulyNatural (Lite), and TrulyNatural STT SDKs differ only in the types of models they support. The APIs, model formats, tools, etc. are identical.

TrulyNatural STT is a strict superset of TrulyNatural (Lite), which in turn is a strict superset of TrulyHandsfree.

TrulyNatural STT

  • Speech-To-Text with transformers and compressed language models.
  • Recognition hypotheses include punctuation and capitalization.
  • Machine-learned NLU for intent and entity identification.
  • Generative language models.
  • Sensory has models available for 35 languages, each in multiple sizes (for best accuracy given a CPU cycle budget). Please contact your account representative or Sensory Sales for details.
  • Includes Open Source software.
  • Features available in TrulyNatural STT only are flagged with stt

TrulyNatural Lite

  • Phonemic acoustic models with FST vocabulary decoding.
  • Grammar-based medium vocabulary command and control.
  • Grammar-based NLU for intent and entity identification.
  • Tools to build recognizers from grammars or phrase lists.
  • API to build or augment recognizers at runtime.
  • Runs VoiceHub "large natural language vocabulary" models.
  • Support for devices with limited RAM (< 1 MiB) and CPU (< 500 MHz).
  • No third-party or Open Source software.
  • Features available in TrulyNatural (Lite and STT) only are flagged with tnl

TrulyHandsfree

  • Fixed, enrolled and adapting wake words.
  • Command sets, which are keyword spotter recognizers for multiple (up to twenty) active phrases.
  • VAD.
  • Command-line tools to enroll and evaluate wake word models, and to convert wake word models into Sensory's THF Micro DSP format.
  • Runs VoiceHub "wake word" and "simple commands" projects.
  • No third-party or Open Source software.

Requirements

For development, you'll need:

Models require audio encoded as 16-bit LPCM and sampled at 16 kHz.

For optimal recognition accuracy, ensure that the dynamic range of the input audio spans at least 12 bits (-24 dBFS peak-to-peak, sample values from -2048 to 2047) and that no clipping is present.

Supported target platforms

TrulyHandsfree and TrulyNatural runs on hundreds of different operating systems and CPU combinations. This distribution includes a subset of these in ~/Sensory/TrulyNaturalSDK/7.6.1/lib/. See the README files in the platform subdirectories for additional details, such as the toolchain and compiler flags used to build the library.

TrulyNatural STT is available for Android, iOS on arm64 and arm64e, macOS, Linux on x86_64, aarch64, and arm, and Windows on x86_64.

Contact us if your target platform isn't listed.

Platform STT support Note
aarch64-linux-gnu yes GLIBC >= 2.33
arm-linux-gnueabi no GLIBC >= 2.17
arm-linux-gnueabihf yes GLIBC >= 2.33
arm-none-eabi no
arm-none-eabihf no
arm-none-eabihf-ethosu no
armv6-linux-gnueabihf no GLIBC >= 2.17
i686-linux-gnu no GLIBC >= 2.17
ios yes 64-bit only
android yes API level >= 21
macos yes
mipsel-buildroot-linux-uclibc no
mipsel-openwrt-linux-musl no
x86_64-linux-gnu yes GLIBC >= 2.17
x86_64-windows-msvc yes Requires MSVC Runtime 2022

Included target platform libraries

Models

TrulyNatural SDK .snsr files include all the models and settings required for a task, and a flow graph that defines the behavior. A task can be a simple as a single-phrase wake word, or something more complicated such as wake word followed by a VAD and an STT recognizer that transcribes the detected speech segment. If you're just interested in the final recognition results, the code required to run these two examples is identical.

This distribution includes sample models and templates used to add additional behaviors to these.

Tools

The ~/Sensory/TrulyNaturalSDK/7.6.1/bin/ directory contains a number of command-line tools. These evaluate models, compose new models, modify settings, enroll wake words, convert wake word models to THF Micro DSP format, and diagnose audio recording quality.

These utilities are compiled for the development host. You can compile these from source for other platforms.

License keys

The TrulyNatural SDK installer embeds the license key entered on the "Product Licensing" page in the libraries and tools it installs. All applications that link against these libraries include this license key.

Keys include the SDK licensee name.

We use license keys to control access to specific SDK features, target platforms, CPU architectures, and to specify an expiration date for access.

Model files also include license keys. These are are validated upon loading.

License keys fall into two broad categories: development ones which either expire at some future date or limit use, and production keys which do not expire and do not have usage limits.

Warning

Do not use development / expiring keys in shipping products. These will stop working when the keys expire.

Contact Sensory to obtain production-ready libraries and models.

license-exp-date, license-exp-message, license-exp-warn, model-license-exp-date, model-license-exp-message, model-license-exp-warn, LICENSE_NOT_VALID, LICENSE_LIMIT_EXCEEDED