1. Presenter: Shih-Hsiang(士翔)
ICASSP 2005

2. Abstract

3. Introduction Speech recognition System can be divided into two parts
Front-End processing (Feature extraction)
Suppress(抑制) the noise
Get more robust parameters
Back-End processing (HMM decoding)
Compensate(補償) for noise
Adapt the parameters
In order to get more noise robust features, there are numerous efforts (based on MFCCs)
Add pre-processing
Noise reduction, speech enhancement
Incorporate algorithms in an MFCC calculation framework
Frequency masking, SNR-Normalization
Add feature post-processing techniques
Cepstral Channel Normalization, Cepstral Mean Normalization

4. Introduction
CMN is known to be a simple noise robust post-feature processing technique
Comparing with cepstral coefficients, the log-energy feature has quite different characteristics
But the log-energy feature (or C0) is treated in the same way as other cepstral coefficients
They propose a log-energy dynamic range Normalization (ERN) method to minimize mismatch between training and testing data

5. Energy Dynamic Range Normalization
Leads to a mismatch
between the clean
and noisy speech
10 dB SNR
Comparing with clean speech, the log-energy feature sequence of noisy speech are
Elevated minimum value
Valleys are buried by additive noise energy, while peaks are not affected as much

6. Energy Dynamic Range Normalization
Log-energy dynamic range of the sequence
Algorithm
Find Max = Max(Log(Energyi) i=1..n )
Min = Min (Log(Energyi) i=1..n )
Calculate target T_Min = αx Max(Log(Energyi) i=1..n )
If Min (Log(Energyi) i=1..n ) < T_Min
For i=1..n 
Liner
Non-Liner

7. Energy Dynamic Range Normalization

8. Experiment Evaluated on the Aurora 2 Relative improvement (R.I.)
Overall accuracy
Tow experiments
Experiment 1：explore how good the performances are in the sense of relative improvement
Experiment 2 ：compare with other techniques

9. Experiment 1

10. Experiment 2

11. Conclusions
The proposed log-energy dynamic range normalization algorithm can have overall about a 30.83% relative performance improvement
It also can be combined with the cepstral mean or variance normalization techniques to achieve an even better result
The proposed does not require any prior knowledge of noise and level
It is difficult to use energy dynamic range normalization to deal with channel distortion
Reducing mismatch in log-energy leads to a large recognition improvement

12. Presenter: Shih-Hsiang(士翔)
ASRU 2003

13. Introduction
Methods of robust speech recognition can be classified into two approaches
Focuses on finding more robust parameters, which are minimally affected by the noise
Formants and their movements (but no effective to estimate)
Focused on compensate for the noise effect
Error rates of automatic could decrease if noise type is known and well trained in the HMM model
But there are countless different kinds of noises in real speech conditions
Training-Testing mismatch always occurs when unknown noise is involved

14. Introduction (cont.)
Human auditory properties and noise masking theory show that spectral peaks (formants) could be used successfully to discriminate speech from the noise
The aim of this paper is to introduce noise reduction and spectral emphasis techniques
It does not require retraining the acoustic models
It utilizes gain coefficients estimated in noise reduction for spectral emphasis

15. Noise Reduction Spectral subtraction
The basic principle is to subtract the magnitude of spectral noise from the noisy speech
Assuming an additive noise model, the input noisy speech signal y(t) is expressed as
time-domain
frequency-domain
Inverse Fourier Transform

16. Noise Reduction & Spectral Emphasis
Thus a noisy signal does not affect the speech signal uniformly over the entire spectrum
Multi-band or non-linear spectral subtraction is better
Over-subtraction factor is frequency dependent
In this paper, gain function is computed as a function of the instantaneous estimated SNRs at each frequency band on a frame-by-frame basis
r : forgetting factor
Noise estimate
b,c : flooring factors
a : scale factor
gain function
Spectral emphasis
m : weight factor

17. Spectral emphasis G(t,f) is an SNR related value
Higher G(t,f) value at the spectral peaks (formants) than that of spectral valleys (buried by noise)
In order to compensate spectral mismatch between clean and noisy speech, some methods have been proposed
peak-to-vally ratio locking, SNR normalization (manipulate values of the spectral valleys)
But this paper is to emphasize values of spectral peaks
It is a trade-off to select a proper weight factor
suggest to use a larger factor to emphasize the formants to discriminate them from noise

18. Voice Activity Detector
Non-speech parts will cause insertion error
deleting non-speech frames will improve the performance
The estimated gain function can be a good detector
when a frame with a flooring value in all frequency bands is set to be a non-speech frame
Leave 10-frame non-speech segment before/after each speech segment, and delete any other non-speech frame

19. Experiment Evaluated on the Aurora 2 Four experiments Noise reduction
Total relative performance improvement of 37.76% can be obtained by the proposed noise reduction method only
Spectral emphasis
A more aggressive factor can cause a bigger gain for noisy speech, but it is also harmful to clean speech
Voice activity detector
It works well especially in low SNR condition (5% improvement)
Noise reduction + Spectral emphasis + Voice activity detector

20. Recognition results – Noise Reduction
From 38.61% to 67.02%

21. Recognition results – Spectral Emphasis

22. Recognition results

23. Recognition results

24. Conclusions
Noise reduction and spectral emphasis techniques are used to improve ASR performance in noisy conditions
The proposed algorithms can easily be embedded in a standard front-end MFCC calculation program
With a low computational load and for real-time operation
It works well for all 8 types of test noises as well as noise plus channel distortion