Tandem Connectionist Feature Extraction for Conversational Speech Recognition

Published on 2007-02-255050 Views

Qifeng Zhu

Multi-Layer Perceptrons (MLPs) can be used in automatic speech recognition in many ways. A particular application of this tool over the last few years has been the Tandem approach, as described by Her

Workshop on Multimodal Interaction 2004 - Martigny

Related categories

Neural Networks

Presentation

Tandem Connectionist Feature Extraction for Conversational Speech Recognition00:01

Using Multi-Layer Perceptron (MLP) in Feature Extraction for Speech Recognition00:23

MLP outputs as features to HMM01:41

*1 Simple and Regular Within-Class Distribution03:04

Exp. 1: Posterior Feature Space04:09

Exp. 2: Log Posterior Feature Space04:55

Exp. 3: Typical Distributions of Log Posteriors in Histogram05:48

*2 Reducing Speaker Variation06:24

Exp. 4: Variances of (Speaker Adaptive Training) SAT Transforms for Different Speakers08:09

*3 Feature Combination: Better Performance, No Dimensionality Increase09:18

Usually What to Expect for a Feature Transform10:23

The Feature Generation Diagram11:06

Some Practical Details in Feature Generation and HMM Decoding12:30

Recognition Experiments13:31

Recognition with a ‘Plain’ System with ML Training14:30

Concerns for a Novel Feature: Scale and Carry Through15:29

Results with Adaptation16:04

Results in a Full-Fledged System16:49

Summary18:26