Abstract
About this talk
Introduction and purpose: Tongue cancer resection affects articulation, which is one of the components of orofacial function. The evaluation of articulation in tongue cancer patients is mainly conducted by speech therapists. However, this evaluation has the drawbacks of requiring significant manpower and involving subjective evaluation.
In this study,1. Using speech recognition software-considered a simple and objective method-, we analyzed changes in articulation before and after surgery.2. We developed a reference model for vowel visualization in healthy individuals using AI.
Furthermore, we attempted to visualize the postoperative vowels of tongue cancer patients.
Subjects and Methods: Patients; 60 patients of tongue cancer treated with surgery were enrolled. 31were male and 29 were female, with an average age of 54.5 (range, 29–81) years. The patients were classified into the closure group (30 patients), who underwent partial tongue resection and primary suture, and the forearm group (22 patients), who underwent tongue resection and reconstruction with a radial forearm flap, and the rectus group (8 patients), who underwent Subtotal mobile tongue component resection and reconstruction with the rectus abdominis flap.
1. Evaluation of speech intelligibility using speech recognition software: UD Talk was used as the speech recognition software. Patients read aloud the story “The North Wind and the Sun,” and UD Talk transcribed the audio into text. The accuracy rate was used as a measure of speech intelligibility and was calculated on a scale ranging from a minimum of 0% to a maximum of 100%. Speech intelligibility was evaluated in each patient preoperatively and 1, 3, 6, and 12 months postoperatively.
2. Development of reference model for vowel visualization in healthy individuals using AI: the vowels were extracted from audio data recorded by 100 professional speakers. The extracted vowels were converted into a 2-dimensional mel-spectrogram. The converted data was fed into the autoencoder, which was trained on the data and converted it into two-dimensional latent variables. A standard model for visualizing vowels, in which each vowel is color-coded, was created.
Results:
1. speech intelligibility: Regarding preoperatively, there was no significant difference among the three groups. Regarding the forearm group, from 1 to 6 months postoperatively, the group showed significantly lower than the closure group, but at 12 months postoperatively, there was no significant difference compared to the closure group. Regarding the rectus group, the rectus group showed significantly lower than the closure group, postoperatively. The speech intelligibility of the rectus group at 12 months postoperatively was classified into the good or poor cases.
2. Reference model for vowel visualization in healthy individuals and visualization of vowels in the rectus group at 12 months postoperatively: We attempted to visualize vowels among our staff. As a result, each vowel was correctly distributed within the same region, respectively. On the other hand, in the rectus group, all cases showed that [e] was recognized as [u]. Furthermore, the poor cases showed that [a] was recognized as [o].
Speaker
About the speaker

