DistillW2V2: A Small and Streaming Wav2vec 2.0 Based ASR Model
read the original abstract
Wav2vec 2.0 (W2V2) has shown impressive performance in automatic speech recognition (ASR). However, the large model size and the non-streaming architecture make it hard to be used under low-resource or streaming scenarios. In this work, we propose a two-stage knowledge distillation method to solve these two problems: the first step is to make the big and non-streaming teacher model smaller, and the second step is to make it streaming. Specially, we adopt the MSE loss for the distillation of hidden layers and the modified LF-MMI loss for the distillation of the prediction layer. Experiments are conducted on Gigaspeech, Librispeech, and an in-house dataset. The results show that the distilled student model (DistillW2V2) we finally get is 8x faster and 12x smaller than the original teacher model. For the 480ms latency setup, the DistillW2V2's relative word error rate (WER) degradation varies from 9% to 23.4% on test sets, which reveals a promising way to extend the W2V2's application scope.
This paper has not been read by Pith yet.
Forward citations
Cited by 2 Pith papers
-
Towards Data-free and Training-free Compression for Speech Foundation Models Using Parameter Clustering
K-means parameter clustering enables data-free training-free pruning of HuBERT and Whisper models with reported WER gains over magnitude pruning on LibriSpeech at 50% and 10% sparsity.
-
Online Predictive Coding for Dual-Mode Self-Supervised Speech Model
Proposes OPC and dual-mode LN to improve dual-mode SSL speech models, reducing WER gap at 160 ms latency on LibriSpeech from 3.65% to 3.40% (test-clean).
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.