Electro-myographic patterns of sub-vocal Speech: Records and classification

Abstract

Abstracthis paper describes the results obtained from recording, processing and classification of words in spoken Spanish by means of analysis of subvocal speech signals. The processed database has six words (forward, backward, right, left, start and stop), In this article, the signals are sensed with surface electrodes (placed on the surface of the throat) and acquired at a sampling frequency of 50 kHz. The signal conditioning consists of a couple of steps, namely the location of area of interest, using energy analysis; and a filtering stage, using Discrete Wavelet Transform. Finally, feature extraction is achieved in the time-frequency domain using Wavelet Packet and statistical techniques for windowing. Classification is carried out with a back propagation neural network whose training is performed with 70% of the database obtained. The correct classification rate was 75%±2.Keywords: Electromyography, subvocal speech, Wavelet, neuronal network. Los autores declaran que no tienen conflicto de interés. TResumen n este trabajo, se describen los resultados obtenidos a partir de los registros, procesamiento y clasificación de las palabras en el idioma español a través del análisis de las señales de voz subvocal. La base de datos procesada tiene seis palabras (adelante, atrás, derecha, izquierda, arranque y parar), en este artículo, las señales fueron detectadas con electrodos superficiales colocados sobre la superficie de la garganta y adquiridas con una frecuencia de muestreo de 50 kHz. El acondicionamiento de la señal consiste en: la ubicación del área de interés mediante el análisis de la energía, y la filtración usando Transformada Discreta Wavelet. Por último, la extracción de características se realiza en el dominio de tiempofrecuencia utilizando Wavelet Packet y técnicas estadísticas para ventanas. La clasificación se llevó a cabo con una red neuronal de retropropagación cuya formación se llevó a cabo con 70% de la base de datos obtenida. La tasa de clasificación correcta fue de 75% ± 2.Palabras clave: Electromiografía, habla subvocal, Wavelet, red neuronal.

Subvocal Speech Recognition via Close-Talk Microphone and Surface Electromyogram Using Deep Learning

“…Wand et al [5] achieved a 34.7% word error rate (WER) on 50 phrases using sEMG signals of 6 facial muscles from 6 subjects. Mendoza et al [6] obtained a WER of 25% from a single sEMG channel but only for 6 Spanish words. Wand et al reported a 54.7% WER on 50 phrases using 35 sEMG channels and 6 subjects [7].…”

“…The reason can be attributed to the nature of the sEMG signal, which is highly variant from subject to subject depending on the muscle strength and gender. Most of the reported research uses facial muscles to capture speech signals [6]- [8]. While giving better accuracy, this placement isn't user friendly and may not lend itself to practical implementations.…”

Silent Communication Device

“…Similar devices have been made as shown in projects such as [23] and [33] in which said projects employ wavelet analysis (explained in much greater detail later in chapter 3) in order to effectively process subvocal speech to allow said subvocal speech to be more easily classified as its intended speech for the final output. The wavelet analysis allows these devices to classify subvocal speech up to accuracies between 70% and 80%, but only through the use of expensive and not pocket-sized equipment.…”

“…As previously mentioned, the motivation before this project was inspired papers such as "Electro-myographic patterns of sub-vocal Speech: Records and classification," by L. E. Mendoza, J. P. Rodríguez, and J. L. R. Valencia [23] and "Sub-vocal Phoneme-Based EMG Pattern Recognition and its application in Diagnosis," by M. Jahan and M.…”

“…Although other means of extracting features from subvocal signals exist such as using Independent Component Analysis instead of wavelet analysis [3] that give much more accurate classification results, these methods are much more complex than that of wavelet analysis and thus would require much more complex programs to implement them, which microcontrollers are unlikely to be able to process without at least very slow or even poor performance, while [31] has proven that wavelet analysis works with microcontrollers effectively. Combining the wavelet analysis through use of a microcontroller [31] (using the more advanced microcontroller MSP432P401R instead of a MSP430 microcontroller), the amplification and filtering circuit of [3], and the subvocal speech translation methods of [23] and [33], the Silent Communication Device could be brought to fruition.…”