Skip to content

Latest commit

 

History

History
66 lines (51 loc) · 3.11 KB

README.md

File metadata and controls

66 lines (51 loc) · 3.11 KB
Raw

Build Status

Enable the DeepSpeech2 in a browser

This repo provides a web application that uses the tensorflow.js to serve the pre-trained DeepSpeech2 model in a browser. It processes the audio input and feeds the audio spectrogram into the model. It also contains the CTCDecoder implementation to process the output of the DeepSpeech2 model, including character level and word level n-gram.

DeepSpeech2 pre-trained model

The model and dataset for the pre-trained model are here:

You can find detailed information about the model and dataset in the links above. We trained the model for 20 epochs and save the model into SavedModel format. In order to use the model in tensorflow.js, we use tfjs-converter to convert the SavedModel to Web friendly format for tensorflow.js. Here is the pre-trained model. The configuration for the model is 4 Bi-directional RNN with 512 nodes. After you download the model, you need to extract the model files into public/model directory. For example:

# Assuming you are at the repository's root directory
mkdir -p public/model

# Change the model.tar.gz to where your download file is
tar xf model.tgz -C public/model

N-Gram language model and Trie

In order to reduce the network overhead, we chose 3-gram ARPA LM and pruned with theshold 1e-7. Its binary format is about 26MB. You can download it from here. Then extract it to public directory. For example:

# Change the 3-gram.binary.tar.gz to where your download file is
 tar xf 3-gram.binary.tar.gz -C public

And trie file is here. Please also extract the trie file to public directory. For example:

# Change the trie.binary.tar.gz to where your download file is
 tar xf trie.binary.tar.gz -C public

Application Set Up

We implement the data pre-process and Connectionist temporal classification(CTC) decoder with N-Gram language model in tensorflow.js. Part of the CTC decoder is implemented in WebAssembly. You can see how to build the WebAssembly module in README under kenlm-wasm directory. Since WebAssembly module is LLVM bitcode, you can also directly use the bitcode under kenlm-wasm directory by directly copy module.js and module.wasm into public/js directory.

Here is the instruction to set up the web application:

  • Run npm install to install dependencies packages
  • Run npm run build to compile the TypeScript source codes

Launch Web Application

Run the web application by running npm run start command. It will start the web server in http://localhost:3000 .

You can also access the demo website here: https://ds2-tfjs.mybluemix.net/