This repository is based on Tensorflow Speech Commands framework (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/speech_commands)
Presented code was used to get into top 17% in Kaggle Speech Recognition Challenge (https://www.kaggle.com/c/tensorflow-speech-recognition-challenge). Architectures that we tried and their corresponding private leaderboard scores are presented in the table below.
| Architecture | Private score |
|---|---|
| wave-net[1] | 0.87137 |
| ds-cnn[2] | 0.84600 |
| cnn on raw audio[3] | 0.83895 |
| lace[4] | 0.83319 |
| crnn[2] | 0.83143 |
| gru-rnn[2] | 0.82978 |
To replicate best result modify pathes and run launch_training_locally.sh. You should have tensorflow of version 1.4 installed.
Example:
#!/usr/bin/env bash
# Path to python
TF_PY="/home/username/tensorflow/bin/python3"
TRAIN_DATA_DIR="/home/username/train_data/audio/"
TEST_DATA_DIR="/home/username/test_data/audio/"
CHECKPOINT_PATH="/home/username/where_to_save_checkpoints/"
SUMMARIES_DIR=""
CHECKPOINT="/home/username/where_to_save_checkpoints/model_name.ckpt-10000"
MODEL_CONFIG="model_configs/best_wave_net.config"
$TF_PY train.py --data_dir=$TRAIN_DATA_DIR --start_checkpoint=$CHECKPOINT --checkpoint_dir=$CHECKPOINT_PATH --arch_config_file=$MODEL_CONFIG --summaries_dir=$SUMMARIES_DIR
References:
1 - Oord A. et al. Wavenet: A generative model for raw audio //arXiv preprint arXiv:1609.03499. – 2016. ↩
2 - Zhang Y. et al. Hello Edge: Keyword Spotting on Microcontrollers //arXiv preprint arXiv:1711.07128. – 2017. ↩
3 - Dai W. et al. Very deep convolutional neural networks for raw waveforms //Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE International Conference on. – IEEE, 2017. – С. 421-425. ↩
4 - Xiong W. et al. The Microsoft Conversational Speech Recognition System //Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE International Conference on. – IEEE, 2017. ↩