Knowledge distillation is an effective machine learning technique to transfer
knowledge from a teacher model to a smaller student model, especially with
unlabeled data. In this paper, we focus on knowledge distillation for the RNN-T
model, which is widely used in state-of-the-art (SoTA) automatic speech
recognition (ASR). Specifically, we compared using soft and hard target
distillation to train large-scaleRNN-T models on the LibriSpeech/LibriLight
public dataset (60k hours) and our in-house data (600k hours). We found that
hard tar-gets are more effective when the teacher and student have different
architecture, such as large teacher and small streaming student. On the other
hand, soft target distillation works better in self-training scenario like
iterative large teacher training. For a large model with0.6B weights, we
achieve a new SoTA word error rate (WER) on LibriSpeech (8% relative
improvement on dev-other) using Noisy Student Training with soft target
distillation. It also allows our production teacher to adapt new data domain
continuously.

本文研究了将知识从一个训练规模较大的教师模型转移到较小的学生模型中的知识蒸馏技术，在 LibriSpeech / LibriLight 公共数据集（60k 小时）和我们公司的内部数据（600k 小时）上对大规模 RNN-T 模型的软目标和硬目标蒸馏进行了比较，发现当教师和学生具有不同的架构（如大教师和小流式学生）时，硬目标更有效。此外，软目标蒸馏在自训练场景（如迭代大型教师训练）中效果更好。通过使用软目标蒸馏进行 Noisy Student 训练，成功在 LibriSpeech 上实现了新的 SoTA 词误率（dev-other 上 8％的相对改进），并允许我们的生产教师不断适应新的数据域。