[2021-12-23] 오늘의 자연어처리

Mixed Precision DNN Qunatization for Overlapped Speech Separation and Recognition

 

Recognition of overlapped speech has been a highly challenging task to date. State-of-the-art multi-channel speech separation system are becoming increasingly complex and expensive for practical applications. To this end, low-bit neural network quantization provides a powerful solution to dramatically reduce their model size. However, current quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different model components to quantization errors. In this paper, novel mixed precision DNN quantization methods are proposed by applying locally variable bit-widths to individual TCN components of a TF masking based multi-channel speech separation system. The optimal local precision settings are automatically learned using three techniques. The first two approaches utilize quantization sensitivity metrics based on either the mean square error (MSE) loss function curvature, or the KL-divergence measured between full precision and quantized separation models. The third approach is based on mixed precision neural architecture search. Experiments conducted on the LRS3-TED corpus simulated overlapped speech data suggest that the proposed mixed precision quantization techniques consistently outperform the uniform precision baseline speech separation systems of comparable bit-widths in terms of SI-SNR and PESQ scores as well as word error rate (WER) reductions up to 2.88% absolute (8% relative).

 

 

 

 

English2Gbe: A multilingual machine translation model for {Fon/Ewe}Gbe

 

Language is an essential factor of emancipation. Unfortunately, most of the more than 2,000 African languages are low-resourced. The community has recently used machine translation to revive and strengthen several African languages. However, the trained models are often bilingual, resulting in a potentially exponential number of models to train and maintain to cover all possible translation directions. Additionally, bilingual models do not leverage the similarity between some of the languages. Consequently, multilingual neural machine translation (NMT) is gaining considerable interest, especially for low-resourced languages. Nevertheless, its adoption by the community is still limited. This paper introduces English2Gbe, a multilingual NMT model capable of translating from English to Ewe or Fon. Using the BLEU, CHRF, and TER scores computed with the Sacrebleu (Post, 2018) package for reproducibility, we show that English2Gbe outperforms bilingual models (English to Ewe and English to Fon) and gives state-of-the-art results on the JW300 benchmark for Fon established by Nekoto et al. (2020). We hope this work will contribute to the massive adoption of Multilingual models inside the community. Our code is made accessible from Github.

 

 

 

 

Toward Educator-focused Automated Scoring Systems for Reading and Writing

 

This paper presents methods for improving automated essay scoring with techniques that address the computational trade-offs of self-attention and document length. To make Automated Essay Scoring (AES) more useful to practitioners, researchers must overcome the challenges of data and label availability, authentic and extended writing, domain scoring, prompt and source variety, and transfer learning. This paper addresses these challenges using neural network models by employing techniques that preserve essay length as an important feature without increasing model training costs. It introduces techniques for minimizing classification loss on ordinal labels using multi-objective learning, capturing semantic information across the entire essay using sentence embeddings to use transformer architecture across arbitrarily long documents, the use of such models for transfer learning, automated hyperparameter generation based on prompt-corpus metadata, and, most importantly, the use of semantic information to provide meaningful insights into student reading through analysis of passage-dependent writing resulting in state-of-the-art results for various essay tasks.