1. III. Analysis of Modulation Metrics IV. Modifications
I. Introduction
III. Analysis of Modulation Metrics
IV. Modifications
Standard Deviation of MER (σMER)
Increasing Envelope Fine Structure
Speech Transmission Index (STI) predicts speech intelligibility in noisy and reverberant environments
Based on change in modulation depth between clean and degraded envelopes in
125 Hz - 8 kHz octave bands
Originally used artificial probe signals, later adapted to use speech
Empirical methods do not require a priori knowledge of degradation condition
Short-time speech-based STI (ssSTI) algorithm developed by Payton & Shrestha [1]
Reliable predictor of intelligibility using windows as short as 1/3 s
Implemented with Envelope Regression (ER) STI method:
MER computed from single speech string degraded by 200 realizations of noise
Yields 200 realizations of MER for each window
Mean and STD of MER computed across realizations
σMER inversely proportional to
Window Length
Octave Band
Adding envelope fine structure (increasing LPF cutoff) reduces σMER
Low octave bands (< 1kHz) and short windows (< 320 ms) benefit most
Greatest reduction when LPF cutoff is slightly higher than twice the octave band upper cutoff f
Spectral Content after squaring
2x band
center frequency
Theoretical Method
ssSTI
160 ms window
320 ms window
80 ms window
ssSTI vs. Theoretical STI
LPF Cutoff for 95th Percentile of σMER ≤ 0.15
Standard Deviation of MER , 50 Hz LPF
Window Length -
250 Hz
1 kHz Band
4 kHz
Octave Band
320 ms
Lowpass Filter Cutoff (Hz)
Applications: Predict Intelligibility
In fluctuating noise environments
Differences between talkers
In a real-time environment -
250 Hz
1 kHz Band
4 kHz
160 ms
Window
884 Hz LPF
309 Hz LPF
110 Hz LPF
Current Study
Focus on 0 dB SNR with stationary speech-shaped Gaussian noise degradation
Aim – Understand & Correct:
Deviation from Theoretical (TH) Method for windows shorter than 1/3 s
Non-zero STI during silence (data points on vertical axis)
80 ms
Time (s)
Silence Detection
Time (s)
II. Short-time STI Computation
Reduces non-zero STI during silence and deviation from TH method for low STI values
If μxk is below threshold relative to long term mean, Mk set to zero
Long term mean approximated with exponential smoothing
Silence Detector Applied to MER
250 Hz Band, 80 ms Window
8 kHz Band
Extract Envelope
Degraded
Clean Mk
aSNRk
TIk
STI t
125 Hz Band
Transmission Index
Apparent
SNR
Octave Banpass Filter f 2
Modulation Metric
Intensity
(Square)
50 Hz Lowpass Filter
y7[m]
x7[m]
y1[m]
x1[m]
Rectangular
Window
Algebraic Expansion of MER
MER can be algebraically expanded to resemble MTH plus error
Dropping k subscript,
Zero if speech & noise uncorrelated
Time (s)
Not zero because z is correlated with speech
Combined Modifications
ob subscript denotes octave band signal
Modified ssSTI vs. Theoretical STI
320 ms
160 ms
80 ms
Broadband Speech
σMER Inversely proportional to variance of clean speech envelope, σx2
Modified ssSTI
1 kHz Octave Band Speech
clipped to ±15 dB
MER Hz Band, 80 ms Window
1 kHz Band Intensity
During Silence
μxk: Mean of clean envelope, xk[m]
μyk : Mean of degraded envelope, yk[m]
μnk : Mean of noise envelope, nk[m] (not accessible in natural environments)
Cxyk: Covariance of clean and degraded envelopes
σxk2: Variance of clean envelope
αk & βk: Octave band weighting & redundancy correction factors, respectively
Theoretical Method
1 kHz Band Intensity Envelope
V. Conclusions
Time (s)
Modifications significantly improve performance of the ssSTI for windows shorter than 320 ms and provide slight improvement for longer windows
For windows shorter than 160 ms, deviation from TH method is significantly reduced but still greater than that of longer windows (e.g. 1 s)
References:
[1] K. L. Payton, M. Shrestha, 2013 “Comparison of a short-time speech-based intelligibility metric to the speech transmission index and intelligibility data” J. Acoust. Soc. Am., 134,
Theoretical Method
(requires noise envelope mean)
Envelope Regression
Method
Reduce σMER → Reduce Deviation from MTH

2. 2aSC7: Error Analysis and Modifications to the Short-Time Speech Transmission Index
Karen L. Payton and Matthew J. Ferreira; Electrical & Computer Engineering Department
University of Massachusetts Dartmouth, N. Dartmouth, MA
I. Introduction
III. Analysis of Modulation Metrics
IV. Modifications