GETS

Guiding EMG-to-Speech Synthesis via Silent Speech Recognition

GETS: Guiding EMG-to-Speech Synthesis via Silent Speech Recognition

Abstract

Electromyography-to-speech (ETS) offers a promising solution for voicing silent speech, yet conventional methods often struggle to achieve high semantic intelligibility. We introduce GETS, a novel framework that establishes a new state-of-the-art (SOTA) with a Word Error Rate (WER) of 11.89% on the silent EMG test set. GETS comprises a diffusion-based generative network conditioned on EMG signals, integrated with a semantic guidance mechanism derived from silent speech recognition (SSR). Our framework not only demonstrates superior semantic intelligibility by effectively disambiguating underspecified EMG inputs but also faithfully preserves the prosodic characteristics—such as temporal alignment and energy contours—inherent in the original EMG signals. This synergistic fusion of acoustic EMG cues and SSR-based semantic guidance establishes a new benchmark for high-fidelity, naturalistic silent speech communication.

GETS Framework

Audio Samples

The audio samples below are synthesized speech from silent EMG signals.

No. Ground Truth Baseline (Gaddy & Klein) Ours
1
 
WER: 33.33%
WER: 0%
his idea was that meteorites might be falling in a heavy shower upon the planet or that a huge volcanic explosion was in progress this idea was that pure ice might be falling in a heavy shower above the planet or that a huge bulk of the explosion was in barcrest his idea was that meteorites might be falling in a heavy shower upon the planet or that a huge volcanic explosion was in progress
2
 
WER: 50%
WER: 0%
henderson he called you saw that shooting star last night enders in tikon do you follow that shooting star last night henderson he called you saw that shooting star last night
3
 
WER: 33.33%
WER: 0%
there was a mouth under the eyes the lipless brim of which quivered and panted and dropped saliva there was some path under the eyes the limpless brim of which quivered and panded and trapped solaifer there was a mouth under the eyes the lipless brim of which quivered and panted and dropped saliva
4
 
WER: 5.88%
WER: 0%
those who have never seen a living martian can scarcely imagine the strange horror of its appearance those who have never seen a living martian can scarcely imagine the strange horror of its parents those who have never seen a living martian can scarcely imagine the strange horror of its appearance
5
 
WER: 55.56%
WER: 0%
besides that there was quite a heap of bicycles but since then there was quite a game of mysegals besides that there was quite a heap of bicycles
6
 
WER: 20%
WER: 0%
it would seem that a number of men or animals had rushed across the lawn it would seem the number of men or hennaples had rushed across the lawn it would seem that a number of men or animals had rushed across the lawn
7
 
WER: 25%
WER: 4.17%
the case appeared to be enormously thick and it was possible that the faint sounds we heard represented a noisy tumult in the interior the case appeared to me enormously thick and it was possible that the faint sounds we heard were presented towards the tumult at the interior the case appeared to me enormously thick and it was possible that the faint sounds we heard represented a noisy tumult in the interior
8
 
WER: 17.39%
WER: 4.35%
i did not know it but that was the last civilised dinner i was to eat for very many strange and terrible days i did not know it but that was the last civilized sitter i was heath for very many strange and terrible days i did not know it but that was the last civilized dinner i was to eat for very many strange and terrible days
9
 
WER: 30%
WER: 5%
by byfleet station we emerged from the pine trees and found the country calm and peaceful under the morning sunlight by my fleet of station we emerged from the pine trees and found the country calm and made full under the burning sunlight by byfleet station we emerged from the pine trees and found the country calm and peaceful on the morning sunlight
10
 
WER: 22.22%
WER: 7.41%
he saw this one pursue a man catch him up in one of its steely tentacles and knock his head against the trunk of a pine tree he saw this one first one man catch him up in one of his steely tickles and to dock his head against the trunk of a pine tree he saw this one pursue a man catching him up in one of its steely tentacles and knocked his head against the trunk of a pine tree
11
 
WER: 40%
WER: 10%
the horse took the bit between his teeth and bolted the horse looked the mid between the steve and bolted the horse took the bit between his teeth and bolted it
12
 
WER: 75%
WER: 12.5%
such things i told myself could not be some of the things i told myself couldnt be me such things i call myself could not be
13
 
WER: 38.46%
WER: 15.38%
some of the refugees were exchanging news with the people on the omnibuses some of the refugees or extinjing news were the people on the enemy musters some of the refugees were exchanging news with the people on the inner benches
14
 
WER: 57.14%
WER: 21.43%
the chances against anything manlike on mars are a million to one he said the gents against the indulging maylight of mars are a millionth of one if he said it changes against anything manlike on mars or a million to one he said
15
 
WER: 66.67%
WER: 33.33%
fearful massacres in the thames valley fearful masters enter the thames of aeneas fearful maskers and the thames valley