Merge pull request #364 from Inhishonor/wake-word-docs

Author: Stypox

README.md

@@ -47,6 +47,10 @@ Currently Dicio answers questions about:
 
 Dicio uses [Vosk](https://github.com/alphacep/vosk-api/) as its speech to text (`STT`) engine. In order to be able to run on every phone small models are employed, weighing `~50MB`. The download from [here](https://alphacephei.com/vosk/models) starts automatically whenever needed, so the app language can be changed seamlessly.
 
+## Wake Word
+
+Dicio uses [OpenWakeWord](https://github.com/dscripka/openWakeWord) for wake word support, and by defaults it listens for the _Hey Dicio_ keyword. If you would like to use a different keyword, you can download other `.tflite` models from [Open Wake Word](https://github.com/dscripka/openWakeWord/releases/tag/v0.5.1) or from [this collection](https://github.com/fwartner/home-assistant-wakewords-collection). Then head to `Settings > Input and output methods > Import custom wake word` and select the `.tflite` model you downloaded. Alternatively, you can train a wake word model for a keyword of your choice by following this [Jupiter Notebook](https://github.com/dscripka/openWakeWord/blob/main/notebooks/automatic_model_training.ipynb) with this [configuration](meta/openwakeword_training_config.yml).
+
 ## Contributing
 
 Dicio's code is **not only here**! The repository with the *compiler for sentences* language files is at [`dicio-sentences-compiler`](https://github.com/Stypox/dicio-sentences-compiler), the *number parser and formatter* is at [`dicio-numbers`](https://github.com/Stypox/dicio-numbers) and the code for evaluating matching algorithms is at [`dicio-evaluation`](https://github.com/Stypox/dicio-evaluation).

meta/openwakeword_training_config.yml

@@ -0,0 +1,36 @@
+augmentation_batch_size: 16
+augmentation_rounds: 1
+background_paths:
+- ./audioset_16k
+- ./fma
+background_paths_duplication_rate:
+- 1
+batch_n_per_class:
+  ACAV100M_sample: 1024
+  adversarial_negative: 50
+  positive: 50
+custom_negative_phrases: []
+false_positive_validation_data_path: validation_set_features.npy
+feature_data_files:
+  ACAV100M_sample: openwakeword_features_ACAV100M_2000_hrs_16bit.npy
+layer_size: 32
+max_negative_weight: 2700
+model_name: hey_dicio
+model_type: dnn
+n_samples: 50000
+n_samples_val: 2000
+output_dir: ./my_custom_model
+piper_sample_generator_path: ./piper-sample-generator
+rir_paths:
+- ./mit_rirs
+steps: 120000
+target_accuracy: 0.7
+target_false_positives_per_hour: 0.2
+target_phrase:
+- hey dicho
+- hey dicio
+- hey dichio
+- hey disio
+- hey dizio
+target_recall: 0.7
+tts_batch_size: 50