Skip to content

Changes

Version 8

8.5.3

  • Fixed issue with license duration and smaller brick sizes

8.5.2

  • Enabled automatic setting of MAX_TOKENS with newer models

8.5.1

  • Fix rare issue with NNPQ divide-by-zero

8.5.0

  • Add optional support for 64-bit THF-Micro.
  • Add support for new version of THF 7.
  • Now rejecting model versions that are too new for this library.

8.4.0

  • Add API to rewind and fast-forward the audio stream, if buffered. Used for wake-to-command use cases.
  • Add support for new version of THF 7.
  • Add ability to potentially support TensorFlow Lite Micro.

8.3.2

Fix issue with THF 7 and improve THF 7 performance.

8.3.1

Fix issue with no audio buffer and depth > 1.

8.3.0

Fix rare issue with correct recognitions getting discarded.

8.2.2

Fix issue with multiple recognitions and depth >= 2.

8.2.0

  • Improve memory usage for SPP (10-20% space improvement).
  • May also improve processing time, depending on caching.
  • Remove need for SensoryProcessRestart after recognition.
  • SV enabled with multiple channels/depth.

8.1.0

  • Remove redundant appStruct_T. Moved remaining members into t2siStruct.
  • Functions that used to take an appStruct_Tnow take a t2siStruct instead.
  • LPSD is now permitted with multiple frames.
  • Fix for issue in reporting endpoints with depth > 1.
  • Fix for issue where very high scoring recognitions could have unexpected scores.

8.0.0

  • Can process multiple channels and multiple frames in the same channel, in one call. See SensoryAllocMulti, SensoryProcessMultiData, and SensoryGetResult. This saves on memory load time for slow memory.
  • SV cannot be used in multi-processing mode.
  • LPSD functions are hidden now, since the user does not need to call them.
  • Audio functions are hidden now (no more appAudio.c) since the user does not need to call them.
  • Audio is now handled internally, but the user can still assign an audio buffer if desired.
  • No need for user to call SensoryFindStartpoint and SensoryFindEndpoint.
  • Most results formerly in t2siStruct are now in RecoResult structure.
  • Header file security: mismatched sensorylib.h and libTHFMicro.a will cause error code ERR_HEADER_VERSION.

Version 7

7.4.2

  • Fix THF 7 buffering.

7.4.1

  • Add THF 7 support for additional accuracy with some models.

7.4.0

  • Add licensing info for THF-Micro™ library and model. See SensoryLibraryLicenseInfo and SensoryModelLicenseInfo.

7.3.1

  • Report recognition status (pending, success, failed, etc.). This enables arbitration between multiple recognizers.

7.3.0

  • SensoryProcessDataSamples Add to process smaller blocks of audio under 15 msec.
  • brickCount is now included in the t2siStruct.
  • Fix issue with endpoint being incorrect if it is not in the audio buffer.
  • Remove deprecated SDET_SDET mode. Use SDET_LPSD for LPSD.
  • Remove unused SensoryDCBlock.

7.2.0

  • 'Wake-to-Command' handling improved.
  • Now SensoryProcessFeatures can be used to speed up processing for multiple recognizers, by using 'already-created' features.

7.1.0

  • A phrase inside another phrase is now permitted (ex. "wake up" and "wake up now").

7.0.0

  • No special handling is needed for post-process rejection (NNPQ).
  • Models with wider context can now be used.