使用生成對抗網路於強健式自動語音辨識的應用

Exploiting Generative Adversarial Network for Robustness Automatic Speech Recognition

在過去幾年中，深度學習技術的發展在許多領域中大放異彩，應用在語音辨識中也一樣表現優異。儘管語音辨識有了大幅度的改進，然而「雜訊」仍然一定程度的干擾語音辨識之準確度。諸如:背景人聲、火車、公車站牌、汽車噪音、餐館背景雜音…以上皆為易影響語音辨識結果的環境噪音。因此語音辨識的強健性技術研究仍扮演著重要角色。過往於強健性技術的研究主要可區分為以特徵為基礎，以及以模型為基礎兩大面向。以特徵為基礎的強健性技術又可分為特徵正規化以及語音訊號增益。本研究主要採用生成對抗網路（Generative Adversarial Network, GAN）以語音訊號增益方式使用在調變頻譜特徵上。我們的目的在於把受到吵雜環境干擾，或被通道效應破壞的語音特徵轉換成接近乾淨環境下錄製之語音特徵，此方法比起原始梅爾倒頻譜係數特徵可以有效的提升辨識率。

In the recent past, deep learning techniques have reached record-breaking performance in a wild variety of applications like automatic speech recognition (ASR). Even though cutting-edge ASR systems evaluated on a few benchmark tasks have already reached human-like performance, they, in reality, are not robust, in the manner that humans are, to disparate types of environmental noise such as babble, train, bus station, car driving, restaurant, and among others. In view of this, this paper embarks on an effort to develop effective enhancement methods, stemming from the so-called generative adversarial networks (GAN), for use in the modulation domain of speech feature vector sequences. A series of experiments conducted on the Aurora-4 database and task seem to demonstrate the practical merits of our methods.