Comparing the Benefit of Synthetic Training Data for Various Automatic Speech Recognition Architectures

Recent publications on automatic-speech-recognition (ASR) have a strong focus\non attention encoder-decoder (AED) architectures which tend to suffer from\nover-fitting in low resource scenarios. One solution to tackle this issue is to\ngenerate synthetic data with a trained text-to-speech system (TTS) if\nadditional text is available. This was successfully applied in many\npublications with AED systems, but only very limited in the context of other\nASR architectures. We investigate the effect of varying pre-processing, the\nspeaker embedding and input encoding of the TTS system w.r.t. the effectiveness\nof the synthesized data for AED-ASR training. Additionally, we also consider\ninternal language model subtraction for the first time, resulting in up to 38%\nrelative improvement. We compare the AED results to a state-of-the-art hybrid\nASR system, a monophone based system using\nconnectionist-temporal-classification (CTC) and a monotonic transducer based\nsystem. We show that for the later systems the addition of synthetic data has\nno relevant effect, but they still outperform the AED systems on\nLibriSpeech-100h. We achieve a final word-error-rate of 3.3%/10.0% with a\nhybrid system on the clean/noisy test-sets, surpassing any previous\nstate-of-the-art systems on Librispeech-100h that do not include unlabeled\naudio data.\n