Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/matlab-deep-learning/deepspeech

This repo provides the pretrained DeepSpeech model in MATLAB. The model is compatible with transfer learning and C/C++ code generation.
https://github.com/matlab-deep-learning/deepspeech

audio deep-learning deployment matlab matlab-deep-learning pretrained-models speech-to-text transfer-learning

Last synced: 3 months ago
JSON representation

This repo provides the pretrained DeepSpeech model in MATLAB. The model is compatible with transfer learning and C/C++ code generation.

Host: GitHub
URL: https://github.com/matlab-deep-learning/deepspeech
Owner: matlab-deep-learning
License: other
Created: 2022-05-27T00:00:56.000Z (over 2 years ago)
Default Branch: main
Last Pushed: 2023-06-08T17:15:10.000Z (over 1 year ago)
Last Synced: 2024-07-27T14:33:16.806Z (4 months ago)
Topics: audio, deep-learning, deployment, matlab, matlab-deep-learning, pretrained-models, speech-to-text, transfer-learning
Language: MATLAB
Homepage: https://www.mathworks.com/products/deep-learning.html
Size: 514 KB
Stars: 7
Watchers: 6
Forks: 4
Open Issues: 1
Metadata Files:
- Readme: README.md
- License: License.txt
- Security: SECURITY.md

Awesome Lists containing this project

MATLAB-Deep-Learning-Model-Hub - deepspeech - deep-learning/deepspeech) | (Speech to Text <a name="Speech2Text"/> / Robotics)

README

# deepspeech

# Speech-to-Text Transcription Using Deep Speech

This repo enables you to load a pretrained Deep Speech model into MATLAB® and perform speech-to-text transcription [1].

![speech2text image](images/helloWorld.png)

**Creator**: MathWorks® Development

## Requirements
- [MATLAB®](http://www.mathworks.com) R2022a or later
- [Deep Learning Toolbox™](https://www.mathworks.com/products/deep-learning.html)
- [Audio Toolbox™](https://www.mathworks.com/products/audio.html)

To accelerate transcription, a GPU and the following toolbox is recommended:

- [Parallel Computing Toolbox™](https://www.mathworks.com/products/parallel-computing.html)

To evaluate word error rate (WER), the following toolbox is recommended:
- [Text Analytics Toolbox™](https://www.mathworks.com/products/text-analytics.html)

## Get Started
Download or clone this repositiory to your machine and open it in MATLAB®.

Run ``deepspeech_inference.mlx`` to perform speech-to-text conversion on a specified audio file. The script plays the audio file to your default sound card and returns the text.

Run ``deepspeech_streaming.mlx`` to perform speech-to-text conversion on streaming audio input.

Run ``deepspeech_deployment.mlx`` to generate plain C code from the speech-to-text system. The script generates a MEX file which you can run from MATLAB to verify results.

Run ``deepspeech_transferlearning.mlx`` to learn how to retrain weights on the network so that speech-to-text performance is optimized for your needs.

The following files are included in the repo:

**Building blocks:**
- ``deepspeechFeatures.m`` - Extract features for DeepSpeech network
- ``deepspeechBuffer.m`` - Buffer features for DeepSpeech network
- ``deepspeech.m`` - Load DeepSpeech speech-to-text network
- ``deepspeechPostprocess.m`` - Decode output from DeepSpeech network

**All-in-one:**
- ``deepspeech2text.m`` - Transcribe speech to text using DeepSpeech

![inference image](images/inference.png)

- ``deepspeech2text_stream.m`` - Transcribe streaming speech to text using DeepSpeech

![streaming inference image](images/inference_stream.png)

## Network Details
The model provided in this example corresponds to the pretrained Deep Speech model provided by [2]. The model was trained using the Fisher, LibriSpeech, Switchboard, and Common Voice English datasets, and approximately 1700 hours of transcribed WAMU (NPR) radio shows explicitly licensed to use as training corpora.

![network image](images/Network.png)

## Metrics and Evaluation
### Accuracy Metrics
[2] reports a 5.97% word error rate (WER) on the LibriSpeech clean test set. The WER corresponds to the acoustic model (included in this repo) **and** a language model, which is not included in this repo.

### Size
The total size of the model is 167 MB.

## License
The license is available in the [License.txt](License.txt) file in this repository.

## References
[1] Hannun, A. "DeepSpeech: Scaling up end-to-end speech recognition", 2014.

[2] https://github.com/mozilla/DeepSpeech/releases/tag/v0.7.1