Noisy Speech Recognition based on Harmonic Noise Model HNM

1
Noisy Speech Recognition based on HarmonicNoise
Model(HNM)

• Xiuyun Shan
• Shengli Li

2
Idea

• Mimic human auditory systems way of recognizing
  noisy speech.
• Preferentially processing the high energy
  components of the speech signal
• Suppressing the weaker parts.
• Noise affects different bands of speech
  differently
• In spectrogram of noisy speech, regions of low
  SNR will be more corrupt than regions of high SNR

3
Question

• What is the high SNR region of the spectrogram
• Voiced part of a speech signal has higher energy
  and has a harmonic structure.
• Explore the harmonicity of voiced speech for
  noisy speech recognition
• Algorithm
• Weighted Harmonic Noise Model

4
What is an HNM

• HNM model represents speech signals as a sum of a
  harmonic and a noise part.
• The harmonic part corresponds to the
  quasi-periodic components.
• The noise part corresponds to non-periodic
  components.

5
HNM

• The harmonic part contains only harmonic
  multiplications of fundamental frequency and is
  modeled as

• fundamental frequency
• total number of harmonics in the signal.
• amplitude coefficient

6
HNM Parameters Estimation

• Pitch
• Using pitch tracking algorithms in the
  literature.
• A Robust Algorithm for Pitch Tracking (RAPT)"
• Amplitude Parameters
• Least-squares solution for the amplitude
  coefficient is

7
HNM

• When HNM is applied to speech signals corrupted
  by additive noise

8
HNM Parameters Estimation

• Harmonic Part

• Noise Part

9
HNM applied to ASR

• Assumption
• Noise and speech are uncorrelated
• Harmonic and noisy components are uncorrelated
• An estimate for clean Mel Spectral component from
  a noisy-corrupted frame of speech

• Mel spectral components of harmonic part
• Mel spectral components of noisy part
• Scaling factors

10
HNM
11
WHNM(Weighted HNM)

• Using scaled version of the harmonic and noisy
  components to obtain an estimate of the
  underlying clean speech signal.
• But how to estimate ?

• Hard to obtain an estimate for
• held constant over all utterances and a range of
  values
• 0.1

12
WHNM-based MFCC Implementation
13
ASR Experiments

• Training dataHW05Traindata
• Test Data HW05Testdata
• Corrupted by additive noise of three kinds with
  SNR from -530dB
• White noise, music noise, vehicle noise
• Features WHNM-based MFCC feature
• Frames labeled as unvoiced are assigned a pitch
  of 150Hz
• 0.1

14
ASR Results
15
Conclusions

• With no assumption about the corrupting noise
• In lower SNR, the model achieve significant
  recognition accuracy.
• In high SNR condition, the algorithm does not
  perform so well as pure MFCC does.