Evaluating Cloud-Based Speech-to-Text Services
This study evaluated the accuracy of cloud-based speech recognition systems developed by Apple, Google, Amazon, and Microsoft. Testing was performed using the TIDIGITS data set, the most popular benchmark for small-vocabulary speech recognition. The results demonstrated that most cloud-based systems had surprisingly low accuracy, with the exception of Microsoft. The low recognition accuracy of these systems suggests that the rapid transition to large-vocabulary speech recognition services may have come at the cost of sacrificing performance on small-vocabulary tasks.