Word Stress Assessment for Computer Aided Language Learning

J. P. Arias, N. B. Yoma, and H. Vivanco

published in Interspeech, 2009

summarized by Davidson, 2011/09/09

back to homepage

• Language and text independent
  
  • No transcript is required
    
  • But requires a reference utterance
    
• The reference and student utterances are compared to decide whether the stress patterns are the same.
  
• F0 contour and frame energy contour are used for the comparison.
  
• DTW algorithm with MFCCs are used to align the two utterances.
  

• 
  

  Figure 1: Block diagram of the proposed system

  
  
  • Both utterances are aligned by DTW with MFCCs.
    
  • F0 and frame energy contours are used to evaluate the trend similarity and to make the stress decision
    
  • xxx
    
• Duration
  
  • Syllable duration vs. Syllable nucleus duration:
    
    

    Syllable duration	Syllable nucleus duration
    Figure 1: Gaussian approximation of duration measures of non-prominent (solid lines) and prominent syllables (dash-dotted lines).

    
    
  • Almost the same recognition rates are obtained --> we can substitute the syllable duration to a rather reliable syllable nucleus duration
    
  • Normalized by ROS (rate of speech, or simply speech rate) (# of phonemes per unit time)
    
• Energy
  
  • Syllable nucleus RMS energy
    
  • Normalized by dividing it over the mean energy of the utterance
    
• Fundamental frequency (F0) contour
  
  • Post-process to smooth out the contour and a final interpolation between voiced regions to obtain a continuous profile
    
• Spectral emphasis
  
  • RMS energy of the signal in 500~2000Hz obtained from a bandpass FIR (finite impulse response) filter
    

• Stress detector
  
  • Past literature showed that the most reliable acoustic correlates of syllable stress are syllable duration and energy
    
  • Stressed syllables exhibit a longer duration and greater energy in the mid-to-high frequency band
    
    

    Prominent syllables	Non-prominent syllables
    Figure 2: Scattered plots of prominent and non-prominent syllables in terms of log-normalized duration and log-normalized spectral emphasis

    
    
  • Overlapping regions exist due to the reason that stress is the only contributin parameters here, but prominence is dependent on both stress and pitch accent.
    
  • Dashed lines represents the decision boundary formed by multivariate Gaussian distributions of the prominent and non-prominent syllables.
    
  • Both acoustic parameters are logged to achieve a better fit to a Gaussian distribution.
    
• Pitch accent detector
  
  • Intonation event method
    
    • Use RFC model (rise/fall/connection) parameters to uniquely describe pitch accent shapes, called the TILT parameter set
      
    • The F0 contour is divided into 25ms frames and data in each frame is linearly interpolated using a least median squares method to obtain robust regression and deletion of outliers
      
    • Classify each frame as R/F/C depending on its gradient
      
    • Merge all adjacent frames into one interval (an intonational event) if they have the same RFC classification
      
    • Possible candidates of pitch accent are the intonational events that exhibits a rise followed by a fall profile, where the rise section is more relevant to prominence.
      
    • Sluijter and van Heuven suggested that pitch accent can be detected by overall syllable energy and pitch variation.
      
    • Pitch variation --> event amplitude (EvAmp) = |rise amplitude| + |fall amplitude| --> EvAmp * EvDur * EvRel where EvDur = event duration and EvRel = relevance of the event along the utterance (???)
      
  • Intonation events
    
    

    Prominent syllables	Non-prominent syllables
    Figure 3: Scattered plots of prominent and non-prominent syllables in terms of log-normalized overall syllable energy and log-normalized intonational event parameters

    
    
  • As before, dashed lines represents the decision boundary formed by multivariate Gaussian distributions of the prominent and non-prominent syllables.
    
• Prominence detector
  
  • Prominence detector = stress detector + pitch accent detector
    
  • Prominence syllable = stress syllable or pitch accented syllable
    
  • Experiment was done on a subset of TIMIT corpus
    
    • Training set: 3637 syllables, 25 speakers
      
    • Test set: 3643 syllables, 26 speakers
      
  • Detection result
    
    

    Detected as Syllable type	Stressed	Pitch Accented	Stressed+ Pitch Accented	None
    Prominent	650	53	280	271
    Non-Prominent	314	41	50	1984

    
    
  • Recognition rate = (650+53+280+1983) / 3643 = 81.44%
    
  • Insertion rate = (314+41+50) / 3643 = 11.12%
    
  • Deletion rate = 271 / 3643 = 7.44%
    

• Performance agrees with the inter-rater agreement rate of around 80% in past research
  
  • 82% for distinguishing high/low/neutral tone, rated by two human raters by inspecting pitch height only
    
  • 77% for distinguishing high/low tone only
    

back to homepage