Enterprise AI Analysis

The proposed framework enhances intelligibility and robustness in SSI-based recognition through a dual-stage approach. First, recurrent ASR architectures are replaced with a transformer-based speech recognition model to effectively capture long-range dependencies and leverage full utterance-level context. Second, a Large Language Model (LLM) is introduced as a post-processing module to refine the transcribed text, correcting grammatical inconsistencies and resolving linguistic ambiguities.

The transformer-based ASR is more robust to noisy, imperfectly generated inputs (like waveforms synthesized from EMG signals) by allowing it to down-weight corrupted frames and emphasize clearer regions through its attention mechanism. The LLM then refines these transcriptions, correcting grammatical errors, restoring incomplete words, and enforcing semantic fluency, using conservative filtering to maintain semantic reliability and avoid over-correction. This integration of contextual modeling at both acoustic and language levels significantly mitigates ambiguities inherent in SSI.

Experimental results demonstrate substantial improvements in intelligibility. Replacing the RNN-based DeepSpeech recognizer with the Transformer model alone reduced the WER to 32.5%, representing a 9.7% relative improvement. Incorporating the LLM-based correction module further reduced WER to 30%, yielding a 16.6% relative improvement compared to baseline. Overall, this dual enhancement achieved a 6% absolute gain over the original 36% baseline WER, marking a significant step towards unambiguous, clear, and fluent transcription in silent speech interfaces.

In terms of runtime, the Transformer-based system is also more efficient. It reduced the average evaluation time per utterance from 1.42 seconds to 0.73 seconds due to the parallelizable nature of self-attention. While the LLM introduces a modest overhead, raising the total to 0.78 seconds, this remains practical given the substantial intelligibility gains, making the system feasible for near real-time applications.

Future work will extend the role of LLMs beyond grammatical correction to encompass style adaptation, personalization, and dialectal variation. This will enable more nuanced and natural speech synthesis. Furthermore, optimization strategies for lightweight and resilient deployment are planned to allow real-time SSI decoding on resource-constrained devices, broadening the applicability in assistive communication and everyday human-machine interaction.