Robustness aspects of active learning for acoustic modeling

We previously proposed [1] an iterative word-selective training method to cost-effectively utilize data preparation resources without compromising system performance. We continue this work and investigate the robustness of our active learning approach with respect to the starting conditions and further propose a stopping criterion that supports our objective to make effective use of transcription effort while minimizing system error. In particular, we demonstrate robustness to seven initial conditions, showing that we can select around 20 hours of training data and achieve a range of error rates between 8.6% and 9.0%, compared to an error rate of 10% when using all 50 hours of the training set. Additionally, we give empirical evidence that our proposed stopping criterion is in general a good predictor of when the minimum error rate is achieved, demonstrated for each of the initial conditions.