Deep Learning-Based Telephony Speech Recognition in the Wild

In this paper, we explore the effectiveness of a variety of Deep Learning-based acoustic models for conversational telephony speech, specifically TDNN, bLSTM and CNN-bLSTM models. We evaluated these models on both research testsets, such as Switchboard and CallHome, as well as recordings from a realworld call-center application. Our best single system, consisting of a single CNN-bLSTM acoustic model, obtained a WER of 5.7% on the Switchboard testset, and in combination with other models a WER of 5.3% was obtained. On the CallHome testset a WER of 10.1% was achieved with model combination. On the test data collected from real-world call-centers, even with model adaptation using application specific data, the WER was significantly higher at 15.0%. We performed an error analysis on the real-world data and highlight the areas where speech recognition still has challenges.