Fig. 1. System architecture for improving intelligibility of dysarthric speech.

Fig. 2. Scatter plots of the stable-point F1 and F2 values for all CVC utterances in the corpus, with the vowel identified by the corresponding vowel symbol, for both the dysarthric (top panel) and non-dysarthric (bottom panel) speaker (color online). (a) Dysarthric speech and (b) non-dysarthric speech.

Fig. 3. Bark-scaled F1 and F2 formant frequency stable points for the 8 vowel classes of the training set with ellipses representing Gaussian components (color online). Ellipses are centered at component means and radii are set to one standard deviation. Dashed ellipses represent source (dysarthric) data, dash-dotted ellipses represent target (normal) data, and solid ellipses represent source data after transformation (mapped data).

Fig. 4. Boxplot of scores using 21 input feature set configurations and 10 test set rotations. The boxes have lines at the lower quartile, median, and upper quartile values. The whiskers are lines extending from each end of the boxes to show the extent of the rest of the data. Outliers are data with values beyond the ends of the whiskers. (a) Scores averaged over rotations and (b) scores averaged of configurations.

Fig. 5. Example illustrating the use of the crossfade strategy versus a straight-line strategy, using a single hypothetical formant trajectory in a consonant-vowel-consonant context. Even though the crossfade strategy (solid line) prevents discontinuities, it introduces unwanted formant movement, which can lead to incorrect vowel identification. The straight-line strategy, equivalent to the target trajectory (dashed line) at all times throughout the vowel, prevents such movement; its drawbacks are potential discontinuities at the vowel boundaries.

Fig. 6. Analysis of dysarthric speech (a) and synthesis of transformed speech (b). The top panels of each subfigure display F0 values. The middle panels display the speech waveform and energy contour. The bottom panels show spectrograms superimposed with formant frequencies F1–F3.

Number of vowel occurrences in the speech corpus

Sets of features used as input to the transformation function

F1, F2, F3 refer to the first, second, and third formant frequency, estimated either by the median of the trajectory of the center 60% of the vowel region, or by the stable point analysis described in the text. Duration represents the vowel region duration. slopeLeft and slopeRight refer to the median first difference in formant frequency value during the first and last 30% of the vowel region. F2rms refers to a 4-tuple of values calculated by the procedure described in the text. F2poly is a 2-tuple of the first and second coefficients of a second degree polynomial fitted to the F2 trajectory over the vowel region. Finally, energy refers to the median energy over the vowel region.

Output target feature values

Formant frequencies are available as generic and individual values. The former were taken from Peterson and Barney’s work (Peterson and Barney, 1952), while the latter were based on formant frequency values estimated from utterances in the speech corpus.

Stimulus conditions and their associated configurations

Intelligibility of stimulus conditions in percent

Values are for 24 non-expert listeners, except for the last column which contains results from 1 expert listener. Standard deviations are given in parentheses in the penultimate column.

Confusion matrix for selected conditions. Target and response vowels are given in the first column and the first row, respectively