A simple speech recognition application for classifying spoken digits. Mostly based on pannous/tensorflow-speech-recognition. Includes scripts for training and inference (which is also used for testing/benchmarking).
Already trained models are included in models/ directory.
Models were trained on following global_variables.py config:
learning_rate = 0.001
width = 20
height = 10
classes = 10
training_iters = 200
training_batch_size = 256
validation_batch_size = 1024
epochs = 20Therefore the training took 4000 steps (training_iters * epochs), resulting in following accuracy:
| type | Vanilla RNN | LSTM | GRU |
|---|---|---|---|
| accuracy* | 35.88% | 91.97% | 95.31% |
*accuracy was calculated using inference.py script on 6400 samples
TensorBoard logs from training included models are also included in this repo, just in case.
Dependencies are included in requirements.txt (with tensorflow-gpu by default)
Project was developed and runs successfully on tensorflow-gpu 1.13.0 with CUDA 10.0 and Python 3.7.3 on Manjaro Budgie 18.0.4 and following dependencies versions:
- numpy>=1.16.3
- librosa>=0.6.3
- tensorflow-gpu>=1.13.0
- tensorboard>=1.13.1
- tflearn>=0.3.2 Older configurations are untested.