Efficiently Train ASR Models that Memorize Less and Perform Better with Per-core Clipping
Lun Wang, Om Thakkar, Zhong Meng, Nicole Rafidi, Rohit Prabhavalkar, Arun Narayanan
TL;DR
The paper investigates unintended memorization risks in large ASR models and the computational burden of DP-SGD. It introduces per-core clipping (PCC), which clips per-core gradients in data-parallel training to curb memorization with negligible overhead. An adaptive variant (APCC) removes extra hyperparameters by using the minimum per-core gradient norm as the bound, and memorization is quantified via the Secret Sharer framework with an exposure metric defined as $\mathrm{exposure}_{\mathcal{M}}(c, \{r_i\}) = \log_2|\{r_i\}| - \log_2\mathrm{rank}_{\mathcal{M}}(c, \{r_i\})$. Empirical results across Conformer LibriSpeech and large voice-search models show PCC reduces memorization and improves WER, offering a scalable privacy-forward path for ASR.
Abstract
Gradient clipping plays a vital role in training large-scale automatic speech recognition (ASR) models. It is typically applied to minibatch gradients to prevent gradient explosion, and to the individual sample gradients to mitigate unintended memorization. This work systematically investigates the impact of a specific granularity of gradient clipping, namely per-core clip-ping (PCC), across training a wide range of ASR models. We empirically demonstrate that PCC can effectively mitigate unintended memorization in ASR models. Surprisingly, we find that PCC positively influences ASR performance metrics, leading to improved convergence rates and reduced word error rates. To avoid tuning the additional hyperparameter introduced by PCC, we further propose a novel variant, adaptive per-core clipping (APCC), for streamlined optimization. Our findings highlight the multifaceted benefits of PCC as a strategy for robust, privacy-forward ASR model training.