Specifically, vocal characteristics such as pitch, voice quality, loudness, and temporal variability are essential mechanisms for conveying emotion, mood, and personality ( Drager et al., 2010 Evitts & Searl, 2006 Fucci et al., 1995 Kangas & Allen, 1990 McCall et al., 1997). When tested on a 2,500-word vocabulary of continuous phrases among speakers with typical voices and speakers with laryngectomy, their recognition models were able to translate sEMG-based subvocal speech to text with a mean word recognition rate (WRR) of over 90% ( Meltzner et al., 2017, 2018).īy design, many SSR systems (among other AAC technologies) have demonstrated the ability to recover lexical content yet struggle to convey the suprasegmental (prosodic) attributes of the content-vital components of natural speech ( Pullin et al., 2017). The authors further developed phoneme-based speech recognition models that utilized many of the methods evolved from ASR applications, including Mel-frequency cepstral coefficients (MFCCs) that were adapted to the sEMG signal spectrum. (2017, 2018) used small sEMG sensors to conform to the complex anatomy of small articulatory muscles of the face and neck, known to provide high-fidelity recordings of these relatively low amplitude sEMG signals during natural speech ( De Luca et al., 2012 Roy et al., 2007). Prior subvocal speech research has attempted to overcome these shortcomings by coupling advancements in sEMG sensor technology with automated speech recognition (ASR) using machine learning techniques. However, the efficacy of such existing technologies has yet to be established efficacy among those with speech or voice impairments and/or remain limited to relatively small (less than 60 isolated words) vocabulary sets that fall short of continuous speech generation ( Betts & Jorgensen, 2005 Jorgensen et al., 2003 Lee, 2008 Manabe & Zhang, 2004 Meltzner et al., 2008).

Numerous groups have worked to advance sEMG-based SSR using signal-based classification methods for recognizing recited numbers, isolated words, or small phrases among speakers with typical voices ( Jorgensen et al., 2003 Jou et al., 2006 Maier-Hein et al., 2005). sEMG-based SSR operates by recording the electrical signals generated by articulatory muscle contractions during speech production using sensors placed over muscles on the face and neck skin surface. Speech recognition using surface electromyography (sEMG) is a particularly promising method, as it can be performed noninvasively while an individual produces subvocal speech. These systems can leverage articulatory motions via electromagnetic articulography (EMA Fagan et al., 2008 Kim et al., 2017, 2018), ultrasound and optical imaging ( Crevier-Buchman et al., 2011 Denby et al., 2010 Hueber et al., 2010), or the neural pathways contributing to speech via implants in the speech motor cortex ( Brumberg et al., 2013 Guenther et al., 2009) or electroencephalographic sensors ( Porbadnigk et al., 2009). SSR refers to the method of receiving the message and may occur through techniques such as lipreading or via sensors to capture different aspects of the speech production and output the recognized message ( Denby et al., 2010). Subvocal speech is a form of communication in which a speaker makes articulatory movements in the absence of glottal excitation and thereby silently mouths their speech. To overcome these limitations, research has explored subvocal speech recognition (SSR) as an alternative communication method. AAC systems may be used to provide an alternative voice source for those who rely on alaryngeal speech however, these technologies can be limited by poor intelligibility and difficulty training (e.g., with esophageal speech Doyle & Eadie, 2005), tissue viability constraints and complicated air valve pressure maintenance (e.g., with tracheoesophageal speech Hillman et al., 2005 Kramp & Dommerich, 2009), and/or the constant use of one's hand as in electrolaryngeal (EL) speech ( Meltzner et al., 2005 Meltzner & Hillman, 2005). Augmentative and alternative communication (AAC) systems enable those with limited speech production capabilities to input physical gestures or text, which may then be visually displayed and/or acoustically synthesized to support communication.

Free speech to text word 2007