Robust and accurate speech recognition systems can only be realized with adequately trained acoustic models. For common languages, state-of-the-art systems are trained on many thousands of hours of speech data and even with large clusters of machines the entire training process can take many weeks. To overcome this development bottleneck, we propose a parallel implementation of Viterbi training optimized for training Hidden-Markov-Model (HMM)-based acoustic models using highly parallel graphics processing units (GPUs). In this paper, we introduce Viterbi training, illustrate its application concurrency characteristics, data working set sizes, and describe the optimizations required for effective throughput on GPU processors. We demonstrate that the acoustic model training process is well-suited for GPUs. Using a single NVIDIA GTX580 GPU our proposed approach is shown to be 94.8× faster than a sequential CPU implementation, enabling a moderately sized acoustic model to be trained on 1000 hours of speech data in under 7 hours. Moreover, we show that our implementation on a two-GPU system can perform 3.3× faster than a standard parallel reference implementation on a high-end 32-core Xeon server at 1/15th the cost. Our GPU-based training platform empowers research groups to rapidly evaluate new ideas and build accurate and robust acoustic models on very large training corpora at nominal cost.