TF_SpeechRecoChallenge

solutions for https://www.kaggle.com/c/tensorflow-speech-recognition-challenge

• based on TensorFlow - exmaples - speech_commands

install

(sudo) pip install -e .

key results

data_dir=~/Data/TF_Speech/speech_commands
test_dir="/home/ftli/Data/TF_Speech/test"
batch_size=48

1. conv LeaderBoard 0.77

• Sainath T N, Parada C. Convolutional neural networks for small-footprint keyword spotting[C]//Sixteenth Annual Conference of the International Speech Communication Association. 2015.

# conv 0.77
bash run.sh --mode train --model conv \
	--data_dir $data_dir --test_dir $test_dir \
  --hparams ""  --train-opts "" \
  --suffix "" --batch_size $batch_size \
  --opt "{ name: Adam, params: {} }" \
  --training_steps  "25000,25000" --skip-infer false \
  --learning_rate "0.01,0.001" --use-gpu true --feature_scaling ''

2. improved-resnet LeaderBoard 0.89 (basic 0.85)

• Tang R, Lin J. Deep Residual Learning for Small-Footprint Keyword Spotting[J]. arXiv preprint arXiv:1710.10361, 2017.

No FeatureScale + BatchNorm (The paper's architecture) + Mfcc 40: LB < 0.1

No FeatureScale + remove BatchNorm + Mfcc 40: LB 0.84

No FeatureScale + Add BN after input + BatchNorm + Mfcc 40: LB 0.86

FeatureScale(centered mean) + BatchNorm + Mfcc 40: LB 0.85

improvements :

• FeatureScale(centered mean, div variance, e.g. cmvn) + BatchNorm + Fbank80: LB 0.89333

  • during the competition, the LB is 0.88

• others: LB 0.88617 during the competition, the LB is 0.88 commit: 922aa0d85

  • Fbank 80
  • dropout
  • feature scale: centered mean (e.g. cmn)
  • First Conv: kernel_size=(3, 10), strides=(1, 4)
  • use MaxPool + AvgPool

# LB 0.886
dropout=0.8
hparams="resnet_filters=45,resnet_type=c,add_batch_norm=True,add_first_batch_norm=False,freeze_first_batch_norm=False"
suffix="15_filters45_BatchNormTypeC_CMN_Dropout${dropout}_Fbank80"

bash run.sh --mode train --model resnet \
	--data_dir $data_dir --test_dir $test_dir \
  --hparams "$hparams"  --train-opts "--dropout_prob $dropout --dct_coefficient_count 80 --feature_type fbank" \
  --suffix "$suffix" --batch_size $batch_size \
  --opt "{ name: Adam, params: {} }" \
  --training_steps  "25000,25000" --skip-infer false \
  --learning_rate "0.01,0.001" --use-gpu true --feature_scaling 'cmn'

3. improved-densenet LeaderBoard 0.88 (basic 0.86)

• Huang G, Liu Z, Weinberger K Q, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017, 1(2): 3.

• improvements (same as resnet - others)

for l in 8;do
	for k in 16;do
		hparams="add_first_batch_norm=True,freeze_first_batch_norm=False,inital_filters=32,dense_blocks=3,num_layers=${l},growth_rate=${k},add_bottleneck_layer=False,theta=1"
		suffix="_F32_DB3L${l}K${k}_CMN"
		echo '========= $suffix ========='
		bash run.sh --mode train --model densenet \
			--data_dir $data_dir --test_dir ${test_dir} \
			--hparams "$hparams" \
			--suffix "$suffix" \
			--opt "{ name: Adam, params: {} }" \
			--training_steps  "10000,10000,10000,10000" --skip-infer false \
			--learning_rate "0.01,0.001,0.0005,0.0001" --use-gpu true --feature_scaling 'cmn' || exit 1
	done
done

4. ensemble LB 0.89745

• spent little time on it.

5. data argument

• tried changing speed / pitch, but got a little achievement.