A comparative large scale study of MLP features for Mandarin ASR

Abstract

MLP based front-ends have shown significant complementary properties to conventional spectral features. As part of the DARPA GALE program, different MLP features were developed for Mandarin ASR. In this paper, all the proposed frontends are compared in systematic manner and we extensively investigate the scalability of these features in terms of the amount of training data (from 100 hours to 1600 hours) and system complexity (maximum likelihood training, SAT, lattice level combination, and discriminative training). Results on 5 hours of evaluation data from the GALE project reveal that the MLP features consistently produce relative improvements in the range of 15% − 23% at the different steps of a multipass system when compared to the conventional short-term spectral based features like MFCC and PLP. The largest improvement is obtained using a hierarchical MLP approach.

Analysis and comparison of recent MLP features for LVCSR systems

“…The cutoff frequency for both filter-banks is approximatively 10Hz. The output of the MRASTA filtering is then processed according to a hierarchy of MLPs progressively moving from high to low modulation frequencies or equivalently from short to long temporal context [7]. The effect of this sequential processing is that the first MLP trained on short temporal context is effective on most of the phonetic classes apart stops and affricatives.…”

“…Because of the large dimension of these time windows, a number of techniques for efficiently encoding the information have been proposed like MRASTA [3], DCT-TRAPS [4], and wLP-TRAPS [5]. The second direction includes a number of heterogeneous techniques that aim at overcoming the pitfalls of the three-layer MLP classifier, including bottleneck architectures [6], hierarchical architectures [7], and multi-stream approaches [8].…”

“…In our previous related work [7], we investigated a subset of these techniques, namely, the MRASTA processing and its hierarchical version in a Mandarin broadcast LVCSR system developed in the framework of the GALE project 1 . This paper aims at complementing that study including other MLP input features (DCT-TRAPS and wLP-TRAPS) as well as Bottleneck architectures in order to cover all the front-ends that have been proposed and integrated into LVCSR systems.…”

“…The study is carried on the same Mandarin Broadcast system described in [7] and we examined the MLP feature performances as stand-alone front ends and in concatenation with spectral features (MFCC). The remainder of this work is organized as follows.…”

Context-dependent MLPs for LVCSR: TANDEM, hybrid or both?

“…Since in the state-of-the-art ASR systems [9,10] MLPs are mainly used in TANDEM approach as features, in this paper our main goal is to compare the two different acoustic modeling techniques not only on MFCC, but also on concatenated cepstral and posterior features. Therefore, evaluating several feature transformation techniques (SAT, LDA) developed for GMM, the study is also extended to context-dependent MLP features.…”