AdaptiveSpeechRecognition

This repository contains the instruction to continuously train IBM speech Recognition model.

IBM speech recognition

IBM offers different API for creating language models, acoustic models, adding resources to each one of them, training, and testing.

Default recognition request:

You may call the IBM S2T API with providing only the Base-language. (i.e en_broadband_english). Also, the default mode has no acoustic (audio) model.

Create a custom language model (English language):

Creating a custom language model gives us the option to add texts to train the model even further. More specifically, you can create a new language model by specifying what base language (English, Arabic, etc) you want. Each language model can have one language only. This takes a couple of parameters. (Your desired name for the model, description of the model, base language of the model). Once you create a new custom lang model, the API will return its ID number.

bash create-new-en-lang-model.sh name description apikey

create a custom language model (any language):

This is same as above, but here you also have the option to specify the language of your lanauge model. list of models : https://cloud.ibm.com/docs/speech-to-text?topic=speech-to-text-models

bash create-new-lang-model.sh name description model apikey

Add text to a custom language model:

To train a custom language model, text data needs to be added to a custom model. Later you can run the training script to make the IBM service train that language model. To add the text to a language model you may use:

bash add-corpus-lang-model.sh textFile.txt languageModelId name apikey

List all custom Language models:

To see the list of all the created language model you may use:

bash list_all_lang_model.sh apikey

Train a custom Language model:

When you added the text file to your language model, you may run the following to start language model training:

bash train-language-model.sh languageModelId apikey

Reset a custom Language model:

When you are not satisfy with your model, you may restart it with the following command

bash reset-lang-model.sh languageModelId apikey

Check custom Language model progress:

You may run the following code to check the status of the specified custom language model: Language model status options: "Pending": The model is not still ready for the training "ready": The model is ready to be trained "Training": The model is being trained "Available": The model is ready to be used

bash check-lang-model-status.sh languageModelId apikey

Removing a custom language model:

If you need to remove a language model use delete-lang-model.sh. You may use it like this

bash delete-lang-model.sh languageModelId apikey

Creating a custom acoustic(audio) model:

By creating a custom acoustic model we can add audio data to train the model more and more. We may create a new acoustic model by specifying the Speech recognition APIK key, and a name for your newly created acoustic model.

bash create-custom-acustic-model.sh name apikey

Add audio to custom acoustic model:

To train a custom acoustic model, audio data needs to be added to a custom acoustic model. Later you can run the training script to make the IBM service train that acoustic model. To add the audio to an acoustic model you may use:

bash add-audio-to-acusticmodel.sh acousticModelId audioFile.zip apiToken**

List custom acoustic models:

To check the list of all available acoustic models you may use:

bash list-all-acust-models.sh apikey

Training a custom acoustic model:

Semi-supervised (Training an acoustic model along with a language model)

If you have the corresponding text of your audio files, you may create a custom language model solely to train an acoustic model. Once you created the temporary custom language model, add the corresponding text of your audio into a text file and append it to the newly created custom language model. Finally use the following command:

bash train-with-both.sh acousticModelId languageModelId apiToken**

Unsupervised (Training an acoustic model alone)

It is also possible to train the acoustic model without having its corresponding texts. To do that, you may use the following command:

bash train-with-both.sh acousticModelId apiToken**

Authentication:

To train your model using both the acoustic and language model, you also need to generate your token using getKey.py This token may expire every couple of minutes.

** apikey is different than the apiToken. You may find the apiToken by running geyKey.py. On the other hand, apikey can be found in your IBM speech recognition service dashboard.