https://github.com/desh2608/dnn-hmm-asr
Hybrid DNN-HMM model for isolated digit recognition
https://github.com/desh2608/dnn-hmm-asr
Last synced: 2 months ago
JSON representation
Hybrid DNN-HMM model for isolated digit recognition
- Host: GitHub
- URL: https://github.com/desh2608/dnn-hmm-asr
- Owner: desh2608
- Created: 2019-04-25T16:25:43.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2020-12-01T12:44:00.000Z (over 4 years ago)
- Last Synced: 2025-03-18T21:54:42.464Z (3 months ago)
- Language: Python
- Size: 12.7 KB
- Stars: 33
- Watchers: 1
- Forks: 7
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Hybrid DNN-HMM model for isolated digit recognition
Python implementation of a hybrid DNN-HMM models for isolated digit recognition.
Forced alignments are obtained from a GMM-HMM model and used to train the DNN.
The DNN is a simple multi-layer perceptron (MLP) implemented using scikit-learn.
### How to run
```
python3 submission.py train test
```
* `train` is the training data
* `test` is the test data
The optional arguments are:
* `--mode`: Type of model (`mlp`, `hmm`). Default: `mlp`
* `--niter`: Number of iterations to train the HMM. Default = 10
* `--nstate`: Number of states in HMM model. Default = 5
* `--nepoch`: Maximum number of epochs for training the MLP. Default=10
* `--lr`: Learning rate for the MLP. Default=0.01
* `--debug`: Uses only top 100 utterances for train and test
### Training data format
I cannot upload the full training and test data (for copyright reasons), but a small sample of the training data can be found at this [Google Drive link](https://drive.google.com/file/d/1NhF7fuX54jau9iXxuitOfm9QRQPHNW2Q/view?usp=sharing). This should help in understanding the format of the data.
### Help
This code is based on a template provided by Shinji Watanabe (Johns Hopkins University), written for a course project.
For assistance, contact `[email protected]`.