There are many factors to consider when choosing audio data for use in training your acoustic models.  First, is there reliable audio speech data that is available that can be used in a beneficial way to create your custom acoustic models?  There are not a lot of avenues available where you can select and license audio that has been appropriately recorded and transcribed in a fashion that is suitable for acoustic model development.  Most speech recognition companies that develop their own acoustic models hold them close and do not open them up to other companies for future development and customization to particular acoustic needs and environments.  Products like Siri, Google, and Nuance to mention a few, collect audio from users, but they do not make that audio available to third parties for use in custom-focused development.

There are a few sources that sell acoustic speech corpora, but these libraries are limited, very costly, and in many cases, do not fit the exact needs of client requirements.  In addition, many of them are extremely limited and also collected over ten years ago presenting issues with current audio properties, such as conversational social media speech, and also, vocal fracking that has been adopted by today’s younger generation.

When possible, MatrixHCI will use audio corpora that is already available as it saves time and resources.  This previously collected and transcribed audio can be used to build “baseline” models that are then continually added to buy collecting audio from your own environment and used to further train and refine your acoustic models.  Audio collected outside of your actual environment is called “out of domain” audio.  Whereas, audio that is collected using your system and by your customers is considered “in domain” audio which is highly preferred after a baseline acoustic model is developed because it contains the actual audio artifacts that your system configuration produces.  It is estimated that 30 hours of “in domain” audio is equivalent to 300 hours of “out of domain” audio.

Baseline acoustic models are initially created with the goal of about 70% recognition accuracy, and these baseline acoustic models are then improved and refined using your “in domain” audio as it is collected and transcribed.  Because of the vacuum of good usable and affordable acoustic audio corpora, MatrixHCI is in the process of creating high quality audio corpora for use in building these baseline models in ways that yield the highest initial accuracy rate.  MatrixHCI will choose, when available, the best existing audio corpora for your projects, and when necessary, will create audio corpora where none exists for your needs from outside sources or from MatrixHCI’s own internal speech corpora library.