Kaggle: TensorFlow Speech Recognition Challenge

V1 Flow

Sample one of valid labels (+ unknown, silence)
Pick one of the clips or...
...If 'silence' picked, generate silence clips from background noise provided
Randomly mix sample with background noise provided, transform pitch/speed/volume
Compute mel-scaled spectrogram
Scale to match mean, std dev with a pre-fit scaler
...
profit!

Output model activations (after softmax) to CSV for multiple training runs/model variations
Generate submission with voting/averaging strategy
Predict same file many times with different transfromations and average/vote result (?, if performance allows)

Name		Name	Last commit message	Last commit date
Latest commit History 78 Commits
docs		docs
v2		v2
.gitignore		.gitignore
Explore Data.ipynb		Explore Data.ipynb
README.md		README.md
_data-generator-test.ipynb		_data-generator-test.ipynb
_predict-test.ipynb		_predict-test.ipynb
data-generator.ipynb		data-generator.ipynb
gen-submission.ipynb		gen-submission.ipynb
gen-train-data.ipynb		gen-train-data.ipynb
lib.ipynb		lib.ipynb
models.ipynb		models.ipynb
predict.ipynb		predict.ipynb
train.ipynb		train.ipynb
vote-all.ipynb		vote-all.ipynb