Cloud-based voice model training with TPUS
CS-21763
Healthcare
Klinikum rechts der Isar
The Chair of Medical Computer Science examines and develops methods for the use of medical data in clinical applications and clinical research and supports scientists and clinicians to be able to use data in networked environments safely with defined data quality.
The Challenge
The implementation of our project would have been unthinkable without the use of cloud resources. Local resources were not available in this form, and the alternative would have meant to put up and manage everything from scratch. The required computing power particularly played a crucial role, especially with reference to the training of voice models.
The Solution
Our goal was to train highly developed voice models, and we used TPUs to develop a comprehensive language model. This general language model forms the basis for specific application tasks that can then be trained on the basis of this model.
The Results
The adaptation of an open source frameworks for use with TPUV4 was crucial for our success, because we were able to achieve surprisingly short training times of only 6.5 days for extensive models and only 1.5 days for smaller models. Such a time frame would have been unrealistic in the local data center and would have taken weeks or months.
The Highlight
The continuous further development of the Google Cloud Platform has also optimized our process. Compared to 2020, for example, we no longer had to build a cluster manually. When generating the TPUs, this was automatically created, which led to a significant increase in stability.