v0.6.0

What's Changed

• Remove model files by @dscripka in #50
• Auto training by @dscripka in #48
• Properly create cv11_test_clips-Folder by @MeinAccount in #56
• Fix typo in README.md by @HepoH3 in #66
• Example updates by @dscripka in #78
• Pyproject and train.py update by @dscripka in #84
• Fix chunksize in utils.py when batch is less than ncpu value by @dlipatov in #98
• Debounce by @dscripka in #116

New Contributors

• @MeinAccount made their first contribution in #56
• @HepoH3 made their first contribution in #66
• @dlipatov made their first contribution in #98

Full Changelog: v0.5.1...v0.6.0

v0.5.1

Fixed a bug with variable input audio sizes for the predict method that would silently drop audio for array sizes not integer multiples of 1280 samples. New behavior supports an arbitrary number of samples for input. Also miscellaneous other small changes.

What's Changed

• Add 'hey_rhasspy' to the list of model paths by @s-knibbs in #42
• Variable input size by @dscripka in #46

New Contributors

• @s-knibbs made their first contribution in #42

Full Changelog: v0.5.0...v0.5.1

v0.5.0

v0.5.0

Added

• A new wakeword model, "hey rhasspy"
• Added support for tflite versions of the melspectrogram model, embedding model, and pre-trained wakeword models
• Added an inference framework argument to allow users to select either ONNX or tflite as the inference framework
• The detect_from_microphone.py example now supports additional arguments and has improved console formatting

Changed

• Made tflite the default inference framework for linux platforms due to improved efficiency, with windows still using ONNX as the default given the lack of pre-built windows WHLs for the tflite runtime (https://pypi.org/project/tflite/)
• Adjusted the default provider arguments for onnx models to avoid warnings (#27)

Removed

v0.4.0

This release of openWakeWord updates the main predict method to allow for audio frames of arbitrary length to be passed in instead of a fixed 1280 samples.

This provides more flexibility during use and increases the efficiency of the melspectrogram calculation, at the cost of reduced detection latency.

v0.3.1

This is a bugfix release of openWakeWord.

• When using custom verifier models, if a verifier model was not provided for every pre-trained model prediction would fail (#18 ). Pre-trained models without verifiers now function normally.

v0.3.0

This release of openWakeWord includes new functionality for training custom "verifier" models that run after the main openWakeWord models and can significantly improve performance by adapting to target speakers and deployment environments. See the documentation for more details, and an example of how to train and use a verifier model in the training tutorial notebook.

v0.2.0

This is the second release of openWakeWord. It includes a new pre-trained model, "hey jarvis", as well as support for SpeexDSP noise suppression and voice activity detection with Silero's excellent VAD model. These are optional features that can be enabled during initialization of an openWakeWord model that can significantly reduce the false-activation rate in noisy environments.

v0.1.1

First release of openWakeword. This release includes the initial versions of the four pre-trained models listed in the main Readme (alexa, hey mycroft, timers, and weather). It also includes pre-built Python wheels for a wrapper of the SpeexDSP noise suppression algorithm.