36 lines
891 B
Markdown
36 lines
891 B
Markdown
# Android models
|
|
|
|
|
|
### EN
|
|
#### android-en-0.3-20201215-jarbas.tar.gz
|
|
Trained using a personal dataset, for not wake-word the google-commands
|
|
dataset, false_activations-20201215.tar.gz, bg_noise-20201215.tar.gz and
|
|
noises-20190814-elt.tar.gz were used
|
|
|
|
Training data contained several speakers from various countries and genders,
|
|
with about 3 samples per speaker, microphone quality and pronounciation varied
|
|
a lot per speaker
|
|
|
|
Added to this 64 samples of my own voice collected with mycroft-core during
|
|
testing of the initial model available at android-20201215-jarbas.tar.gz
|
|
|
|
|
|
=== TrainData ===
|
|
- wake_words=114
|
|
- not_wake_words=14444
|
|
- test_wake_words=38
|
|
- test_not_wake_words=14496
|
|
|
|
|
|
=== Test set results ===
|
|
- 14513 out of 14533
|
|
- 99.86%
|
|
- 0.07% false positives
|
|
- 26.32% false negatives
|
|
|
|
|
|
- False Positives: 10
|
|
- True Negatives: 14485
|
|
- False Negatives: 10
|
|
- True Positives: 28
|