4-bit Quantization of LSTM-based Speech Recognition Models

Abstract

We investigate the impact of aggressive low-precision representations of weights and activations in two families of large LSTM-based architectures for Automatic Speech Recognition (ASR): hybrid Deep Bidirectional LSTM -Hidden Markov Models (DBLSTM-HMMs) and Recurrent Neural Network -Transducers (RNN-Ts). Using a 4-bit integer representation, a naïve quantization approach applied to the LSTM portion of these models results in significant Word Error Rate (WER) degradation. On the other hand, we show that minimal accuracy loss is achievable with an appropriate choice of quantizers and initializations. In particular, we customize quantization schemes depending on the local properties of the network, improving recognition performance while limiting computational time. We demonstrate our solution on the Switchboard (SWB) and CallHome (CH) test sets of the NIST Hub5-2000 evaluation. DBLSTM-HMMs trained with 300 or 2000 hours of SWB data achieves <0.5% and <1% average WER degradation, respectively. On the more challenging RNN-T models, our quantization strategy limits degradation in 4-bit inference to 1.3%.

Mix-GEMM: An efficient HW-SW Architecture for Mixed-Precision Quantized Deep Neural Networks Inference on Edge Devices

“…However, Mix-GEMM can be applied to all the DNNs quantized with any uniform affine quantization technique, and as such, any advancement in that area can be potentially leveraged by Mix-GEMM. For example, recent works [24], [65], [69] have demonstrated competitive quality of results for low mixed-precision quantization of BERT for Natural Language Processing (NLP), whose compute expansive kernels based on matrix-matrix multiplications could be accelerated exploiting Mix-GEMM.…”

4-bit Conformer with Native Quantization Aware Training for Speech Recognition

“…A standard design approach to fit the ASR model under budget is to apply model quantization or network pruning to large models. Recent arXiv:2312.08553v3 [eess.AS] 16 Jan 2024 quantization studies [22,23,24,25] have shown that it is possible to quantize the ASR models to 4-bit and even 2-bit with only marginal performance loss. Similarly, in terms of network pruning, both unstructured and structured sparsity [17,18,19,20,21] patterns have seen reasonable performance at high sparsity level, through various algorithms based on iterative magnitude pruning.…”

“…From prior studies, we have seen success on end-to-end ASR compression through sparse network pruning [17,18,19,20,21] and model quantization [22,23,24,25] However, compressing these massive universal speech models can lead to new challenges on the top of regular end-to-end models. For example, USMs have much large model sizes, and therefore higher compression ratios are needed to reach the efficiency requirements for deployments.…”

Sub-8-bit quantization for on-device speech recognition: a regularization-free approach

“…Existing quantization methods can be post-training quantization (PTQ) or in-training / quantization aware training (QAT). PTQ is applied after the model training is complete by compressing models into 8-bit representations and is relatively well supported by various libraries [3,4,5,6,7,8], such as TensorFlow Lite [9] and AIMET [10] for on-device deployment. However, almost no existing PTQ supports customized quantization configurations to compress machine learning (ML) layers and kernels into sub-8-bit (S8B) regimes [11].…”