A noninvasive system that converts muscle activity into synthesized speech may help restore intelligible communication following total laryngectomy.
In the pilot study, researchers evaluated a 60-year-old male patient who had undergone total laryngectomy about 2 months prior to testing and a 52-year-old healthy male reference speaker. The patient had previously received radiation therapy for laryngeal cancer, which prevented electrolarynx use because of tissue damage. His tracheoesophageal voice prosthesis had been removed following leakage, aspiration, and infections, leaving typing as his only means of linguistic communication. Both participants silently articulated 499 date-and-time phrases while surface electromyography (EMG) signals were recorded using 20 electrodes positioned on the face, jaw, and neck.
The evaluation focused on speech intelligibility, measured by word error rate based on transcriptions from 5 native English-speaking listeners aged 23 to 25 years, agreement among listeners, the feasibility of reproducing the patient's presurgical voice, and whether a reduced electrode array could capture sufficient muscle activity for speech decoding.
The system generated intelligible speech in both participants, with a mean word error rate of 20% in the patient who underwent laryngectomy vs. 15% in the healthy reference speaker and an 80% and 85% correct word recognition, respectively. The higher error rate in the patient was consistent with the additional challenges of decoding postsurgical compensatory muscle activity.
The researchers demonstrated that synthesized speech could be reproduced in the patient's presurgical voice using about 8 minutes of previously recorded audio. The system converted muscle signals into speaker-independent speech representations before generating audible output, allowing the core decoding process to function without recordings of the patient's original voice. Personalization was performed separately when presurgical recordings were available.
The reported word error rates reflected the synthesized speech before voice personalization. The personalized output was demonstrated qualitatively rather than assessed through a separate quantitative intelligibility measure. The personalization process incorporated an intermediate text-extraction step, which the researchers identified as an area for future development.
Electrode-selection analyses showed that 9 of 20 electrodes in the patient and 11 of 18 usable electrodes in the reference speaker were sufficient to recover 95% of the multichannel signal information. The findings suggested that smaller electrode arrays could support more compact hardware. However, the researchers emphasized that electrode selection was specific to each participant and should not be interpreted as a generalizable placement configuration.
Agreement among listeners was stronger for the healthy reference speaker compared with the patient. Single-trial recordings demonstrated coordinated muscle activity corresponding to silently articulated words, indicating that speech-related muscle patterns remained detectable despite the absence of phonation.
The study was limited by its inclusion of just 2 participants. Testing involved a restricted vocabulary of dates and times rather than spontaneous conversation, and recordings were collected during a single session. The small listener panel and lack of longitudinal evaluation left questions about decoder stability across days and repeated electrode placement. The intelligibility of the personalized voice output was not quantitatively assessed, limiting conclusions about its performance in routine communication.
The findings demonstrated the feasibility of noninvasive speech synthesis following total laryngectomy, including the possibility of reproducing a patient's original voice when presurgical recordings are available. However, larger studies involving more clinically diverse patients, open conversational speech, and repeated testing are needed to establish broader applicability.
“This study serves as an initial step toward understanding and harnessing [surface electromyography] for speech synthesis after total laryngectomy,” wrote lead study author Naresh K. Kaushal, of the Center for Mind and Brain at the University of California, Davis, and colleagues.
The study authors declared no conflicts of interest.