The speech neuroprosthesis

Angrick et al. and Anumanchipalli et al. used a two-stage decoding process in which a deep-learning model regresses neural activity to acoustic features and a vocoder converts them to audible speech.

“Another approach, pursued by Angrick et al.111 and Anumanchipalli et al.110, involved a two-stage decoding process. In the first stage, a deep-learning model is trained to regress neural activity (recorded from SEEG111 or ECoG110) to a time series of acoustic features, such as the mel-spectrogram (Fig. 3b). In the second stage, a speech synthesizer — a vocoder — is used to convert the acoustic rep”

• cites [111] sourcePartly supported

  The Decoding Approach section describes the claimed two-stage process: a DenseNet maps neural activity to mel-scaled spectrograms, and a WaveNet vocoder produces audio. However, the citing sentence assigns SEEG to Angrick et al.; the Experiment and Data Recording section says they recorded ECoG using a cortical electrode grid. Page numbers are unavailable in our copy. (LLM, conf=0.90)

  Experimental — accuracy not yet measured
• cites [110] sourceDisputed

  The Speech decoder design and Methods (Neural decoder; page numbers unavailable) describe a different two-stage process: an ECoG-based neural network first decodes articulatory features, then another network converts those features to acoustic features. Audio is synthesized afterward, not as the second decoding stage. The paper treats direct neural-to-acoustic decoding as a separate comparator. (LLM, conf=0.89)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Imagined, silently attempted and attempted speech reflect intended speech from the participant, distinct from internal monologues or thoughts.

“Together, imagined, silently attempted and attempted speech reflect intended speech from the participant, which is distinct from internal monologues or thoughts139.”

• cites [139] sourceDisputed

  In “Speech Neuroprosthesis as a Mind-Reading Device” (page unavailable), the paper says covert speech can include internal monologue and that distinguishing speech intended for externalisation from speech not intended for it is a central difficulty. It also warns that imagined speech could be captured without the user intending to express it. This contradicts the claim that imagined speech is necessarily intended speech distinct from thoughts or internal monologue. (LLM, conf=0.86)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Studies on cortical encoding of speech features laid groundwork for a speech neuroprosthesis that decoded text from cortical activity of an individual with incomplete locked-in syndrome who relied on AAC.

“Studies furthering our basic understanding of the cortical encoding of speech features, notably motor control of the vocal-tract articulators during speech, laid the groundwork for a speech neuroprosthesis that decoded text in the form of words and sentences from the cortical activity of an individual with incomplete locked-in syndrome who relied on AAC methods to communicate12.”

• cites [12] sourceSupported

Bouchard et al. first demonstrated the dynamics of somatotopic speech-articulator representations in the SMC for the jaw, lips, tongue and larynx during speech production of syllables.

“Bouchard et al.17 first demonstrated the dynamics of somatotopic speech-articulator representations in the SMC for the jaw, lips, tongue and larynx during speech production of syllables (Fig. 1; Bouchard et al. 2013).”

• cites [17] sourceSupported

Neural populations in vSMC and midPrCG control millisecond-level movements of vocal-tract articulators that give rise to speech sounds.

“Neural populations in the ventral sensorimotor cortex (vSMC) and the middle precentral gyrus (midPrCG) control the millisecond-level movements of the vocal-tract articulators (lips, tongue, jaw and larynx) that give rise to speech sounds17,18 (Fig. 2a).”

• cites [17] sourcePartly supported

  The Abstract and Discussion describe vSMC representations of vocal-tract articulators and their coordination in producing speech sounds. The Timing of Articulator Representations section places articulator movements on a scale of tens of milliseconds. The paper does not identify the middle precentral gyrus (midPrCG) as a separate region controlling those movements; page numbers are unavailable in our copy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [18] sourcePartly supported

  The Abstract (page unavailable) reports somatomotor maps of vocal-tract articulation spanning primary motor and somatosensory areas, and the Results locate articulation-related activity on the precentral gyrus. But the paper does not identify the middle precentral gyrus specifically or establish millisecond-level control by these neural populations. The Discussion (page unavailable) says its findings do not directly address the planning and sequencing of articulator movements. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

At the neuronal level, locations on the anterior precentral gyrus in orofacial or hand areas contain single neurons tuned to movements of the whole body, including key speech articulators.

“Interestingly, at the neuronal level, locations on the anterior precentral gyrus in orofacial or hand areas contain single neurons tuned to movements of the whole body, including key speech articulators15,51,52.”

• cites [15] sourcePartly supported

  The Main section (page not provided) reports intermixed tuning to the jaw, larynx, lips and tongue at a single-neuron level in area 6v. It studied attempted orofacial movements and speech, but does not establish tuning to movements of the whole body or report such tuning in hand areas. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [51] sourceSupported
• cites [52] sourcePartly supported

  The Discussion reports neurons responding to mouth and tongue movements in the dorsal arm-and-hand area of motor cortex. It describes representation of much of the body as a suggestion based partly on emerging work, not a finding that individual neurons are tuned to whole-body movements. This study recorded in the hand area of two people with tetraplegia, not in orofacial areas. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Current speech-decoding applications from brain activity are primarily aimed at restoring communication to individuals with both limb and vocal-tract paralysis.

“Current speech-decoding applications from brain activity are primarily aimed at restoring communication to individuals with both limb and vocal-tract paralysis14–16.”

• cites [14] sourceSupported
• cites [15] sourcePartly supported

  The Abstract and Discussion describe speech BCIs as a way to restore communication to people with paralysis and discuss people with severe speech and motor impairment. They do not establish that current applications are primarily aimed at people with both limb and vocal-tract paralysis. The Main section says this participant retained limited orofacial movement and could vocalize, though she could not speak intelligibly. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  The Introduction describes speech brain-computer interfaces as a way to restore lost communication. But the Methods (Study participant) describe a person with severe limb impairment and dysarthria, and the Limitations (Decoding speech in patients with other neurological impairments) state that he retained voluntary bulbar muscle control. The paper therefore does not establish that current applications are primarily aimed at people with both limb and vocal-tract paralysis. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Years of work mapping premotor and primary motor representations revealed a somatotopic organization in which neural control of muscle groups is represented along the ventral-dorsal axis.

“Years of work focused on mapping premotor and primary motor representations on the precentral gyrus and central sulcus, revealing a somatotopic organization in which neural control of muscle groups is represented along the ventral–dorsal axis34–38.”

• cites [34] sourceNot checked — no full text of the source
• cites [35]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [36] sourceNot checked — no full text of the source
• cites [37] sourceNot checked — no full text of the source
• cites [38] sourceSupported

Subsequent work discovered that SMC populations encode articulatory kinematic trajectories and vocal-tract configurations during continuous speech.

“Subsequent work discovered that SMC populations encode articulatory kinematic trajectories and vocal-tract configurations43 during continuous speech24,39,44,45 (Fig. 1; Mugler et al. 2018, Chartier et al. 2018 and Dichter et al. 2018).”

• cites [43] sourcePartly supported

  The Discussion reports that precentral-gyrus activity represented articulatory gestures more strongly than phonemes. But the Results describe recordings during single-word, monosyllabic utterances, not continuous speech. The paper estimated articulator trajectories to label gestures; it does not establish that the recorded populations encoded continuous kinematic trajectories or vocal-tract configurations. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [24] sourceSupported
• cites [39] sourcePartly supported

  The Volumetric Results and Surface Results (pages not provided) locate somatomotor activity associated with articulation, tongue movement, and vocalization during a short syllable task. They do not show that SMC populations encode vocal-tract configurations or articulatory kinematic trajectories during continuous speech. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [44] sourcePartly supported

  The Results section ('dLMC and vLMC encoding of pitch components') reports that sensorimotor-cortex electrodes encode pitch accent, phrase, and voicing during spoken sentences. But the Methods ('Acoustic analysis') explicitly says the model does not track articulators directly or model their movements, so this paper does not establish encoding of articulatory kinematic trajectories. The citing sentence places a separate citation immediately after 'vocal-tract configurations.' (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [45] sourceCould not verify

  The Cerebellum section describes mapping lip, tongue and laryngeal activity during vocalization, and the Motor Cortex section discusses laryngeal representations. Neither reports that SMC populations encode articulatory trajectories or vocal-tract configurations during continuous speech (searched for those terms and related descriptions). The indexed passages provide no page numbers. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

Chartier et al. discovered that the SMC encodes a large variety of low-dimensional movement trajectories of the vocal tract (gestures) that could be directly related to all consonants and vowels.

“Chartier et al.24 discovered that the SMC encodes a large variety of low-dimensional movement trajectories of the vocal tract (also called gestures), which could be directly related to all consonants and vowels.”

• cites [24] sourcePartly supported

  The Results (Kinematic organization of vSMC; page unavailable) report distinct articulatory trajectories associated with broad consonant groupings and vowel clusters, supporting the central claim that vSMC encodes vocal-tract movements related to speech sounds. The paper says the consonant groupings broadly cover English consonants, but its analysis excludes liquids and does not establish a direct correspondence to every consonant and vowel. The Methods describe concise seven-value representations of the trajectories, though they do not characterize the full trajectories as low-dimensional. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Recent studies achieved high-performance speech-decoding results by measuring relevant single-unit and multi-unit neuronal activity from multiple small cortical regions on the anterior precentral gyrus.

“Recent studies have also achieved high-performance speech-decoding results15,16 by measuring relevant single-unit and multi-unit neuronal activity from multiple small (4 × 4 mm2 (ref. 83)) cortical regions on the anterior precentral gyrus15,16,52,98.”

• cites [15] sourcePartly supported

  The Main text reports speech-related recordings from two small area 6v arrays and says both contributed to decoding; the Discussion reports decoding unconstrained sentences at 62 words per minute using neural activity recorded at single-neuron resolution. The passages do not establish that the decoder measured multi-unit activity, so that part of the attribution is unsupported. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  The Methods (Surgical implantation; Recording array implant locations and decoding contributions, pages not provided) describe four small arrays targeting precentral-gyrus regions, with electrodes designed to record from one or a few neurons. The Results (Online decoding performance, page not provided) report high-accuracy speech decoding. The copy does not establish that all the recorded regions were on the anterior portion of the precentral gyrus; it includes an array targeting primary motor cortex (area 4). (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [52] sourcePartly supported

  The Results report above-chance offline decoding of spoken syllables and words from intracortical activity, and describe single-unit and threshold-crossing activity recorded from two arrays per participant in the dorsal motor-cortex ‘hand knob’ area. The Neural recording methods likewise place the arrays in that area, but do not establish the claim’s more specific ‘anterior precentral gyrus’ location or array dimensions; the Discussion also limits the result to a small set of sounds in participants able to speak. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [98] sourcePartly supported

  The Methods describe two small Utah arrays in the dorsal precentral ‘hand knob’ area, and the Discussion reports speech-decoding performance comparable to ECoG studies. But the Methods say the spikes were not sorted into single units, and the Results say high-frequency local field potentials—not spike activity—gave the better decoding performance. The paper does not establish the claim’s specific anterior location or its characterization of the results as achieved by measuring single- and multi-unit activity. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

When able speakers imagine speaking or hearing a word, neural activity in the STG can be used to decode those words with above-chance accuracy.

“Along these lines, when able speakers are instructed to imagine speaking or hearing a word, neural activity in the STG can be used to decode those words with above-chance accuracy59–63.”

• cites [59] sourceSupported
• cites [60] sourcePartly supported

  The Results (§3.1; page unavailable) report above-chance decoding of vowels and consonant pairs in imagined words, and §3.2 identifies temporal speech regions among the informative sites. The study did not demonstrate decoding whole words from STG activity, and its task was covert repetition of visually presented words, not a separate instruction to imagine hearing them. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [61] sourcePartly supported

  The Results (Covert speech; page not provided) report above-chance identification of speech segments during silent reading, and the Discussion identifies the STG as an informative region for decoding covert speech features. The study did not demonstrate decoding individual imagined words or separately test instructions to imagine hearing a word; its identification result concerns segments of continuous speech. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [62] sourcePartly supported

  The Results (‘Decoding imagined speech’; page unavailable) report above-chance classification in studies whose participants imagined hearing or speaking, but the classifier distinguished groups of words rather than decoding individual words. The Results (‘Speech task and item discrimination’; page unavailable) find item-related signals in bilateral superior temporal cortex, but do not establish that STG activity alone decoded the words above chance. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [63] sourcePartly supported

  The Searchlight Results (p. 5) identify STG voxels that distinguish semantic categories when participants were presented with spoken or written words. But the Experimental Design (p. 3) describes covert recall of category exemplars, not instructions to imagine speaking or hearing a word; the paper reports decoding animal-versus-tool categories, not individual imagined words. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

In able speakers, single-word classification has been highly successful using predefined and restricted vocabularies.

“In able speakers, single-word classification has been highly successful using predefined and restricted vocabularies115,116.”

• cites [115] sourceSupported
• cites [116] sourcePartly supported

  The “Decoding inner speech using electrocorticography” section reports significantly above-chance classification of individual words across five patients, with a mean accuracy of 69% for word pairs versus 50% chance. The “Opportunities” section describes this as decoding a limited vocabulary. Calling it “highly successful” overstates what the paper presents as a proof of concept, and the cited passages do not establish that the patients were able speakers (page numbers unavailable in our copy). (LLM, conf=0.83)

  Experimental — accuracy not yet measured

When an individual with anarthria silently attempted to speak, select neural populations in the STG contributed to decoding performance.

“Furthermore, when an individual with anarthria silently attempted to speak — that is, was asked to attempt to ‘mime’ with no audible output from that attempt — select neural populations in the STG contributed to decoding performance14.”

• cites [14] sourcePartly supported

  The Overview of multimodal speech-decoding system says the participant had anarthria, silently attempted speech without vocalizing, and had electrodes over the superior temporal gyrus (STG). Articulatory representations drive decoding says temporal-lobe electrodes contributed to decoding, but cautions that they may have recorded activity from the sub-central gyrus rather than temporal-lobe sites. It therefore does not establish that neural populations in the STG itself contributed. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Makin et al. applied an RNN encoder-decoder framework to encode temporal patterns in neural activity during sentence production into an abstract representation that was then decoded word-by-word into phrases.

“To build on this result and move beyond direct classification of words or sentences, Makin et al.108 applied a recurrent neural network (RNN) encoder-decoder framework to encode temporal patterns in neural activity during sentence production into an abstract representation that was then decoded word-by-word into phrases (Fig. 1; Makin et al. 2020).”

• cites [108] sourceSupported

Intracortical MEA recordings measure neural activity at the level of single neurons or neuron clusters.

“Intracortical MEA recordings are used to measure neural activity at the level of single neurons or neuron clusters83.”

• cites [83] sourceSupported

Semantics appear to be encoded in a distributed manner across many cortical regions including prefrontal, speech and language association cortices.

“Semantics appear to be encoded in a distributed manner, across many cortical regions including prefrontal, speech and language association cortices67,68 (Fig. 1; Tang et al. 2023).”

• cites [67] sourceSupported
  Experimental — accuracy not yet measured
• cites [68] sourceSupported
  Experimental — accuracy not yet measured

Durable stability has enabled ECoG-based neuroprostheses without the need for daily calibration by the user.

“Despite this, durable stability has enabled ECoG-based neuroprostheses without the need for daily calibration by the user12,14,94,95.”

• cites [12] sourceSupported
• cites [14] sourceSupported
• cites [94] sourceNot checked — no full text of the source
• cites [95] sourceSupported

Demonstrations in able speakers with epilepsy served as proof-of-principle for future applications in individuals with vocal-tract paralysis that led to the first clinical trial for a speech neuroprosthesis using a chronically implanted 128-electrode ECoG array.

“Importantly, those demonstrations in able speakers with epilepsy served as a proof-of-principle for future applications in individuals with vocal-tract paralysis that led to the first clinical trial for a speech neuroprosthesis using a chronically implanted 128-electrode ECoG array12.”

• cites [12] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

A 253-electrode ECoG array with increased channel count and density led to substantial gains towards high-performance speech decoding.

“Furthermore, a 253-electrode ECoG array with increased channel count and density led to substantial gains towards high-performance speech decoding14.”

• cites [14] sourceSupported

Three studies in individuals with vocal-tract paralysis used RNN models trained with CTC loss to map neural activity sequences to phoneme sequences without needing alignment.

“Three studies (each in a different individual with vocal-tract paralysis: two with MEAs15,16 and one with ECoG14) leveraged RNN models trained to map an input sequence of neural activity to an output sequence of phonemes without the need for alignment between the neural activity and ground-truth sequence of phonemes”

• cites [15] sourceSupported
• cites [16] sourcePartly supported

  The Methods (‘Decoding speech’; page unavailable) describe predicting phonemes from cortical neural activity during attempted speech, and the Introduction identifies the participant as having advanced ALS and severe dysarthria. Our copy does not establish that this study used an RNN trained with CTC loss or decoded without phoneme alignment; it also describes impaired speech, not complete vocal-tract paralysis. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [14] sourceSupported

Development of a spelling system in which the same participant with anarthria attempted to silently speak 26 NATO code words facilitated access to a larger vocabulary.

“The development of a spelling system in which the same participant with anarthria attempted to silently speak 26 NATO code words — phonetically discriminable words which represent each letter in the alphabet — rather than the 50 common words previously used facilitated access to a larger vocabulary120.”

• cites [120] sourceSupported

Herff et al. demonstrated that during continuous speech, phoneme sequences could be decoded provided synchronized produced acoustics were available to aid model training.

“Herff et al.106 further demonstrated that, during continuous speech, phoneme sequences could be decoded provided that synchronized produced acoustics were available to aid model training106.”

• cites [106] sourceSupported

The concatenative approach built a brain-to-speech lookup library by associating 150 ms neural activity segments with synchronized 150 ms audible speech.

“They first built a brain-to-speech lookup library during training by associating 150 ms neural activity segments with the synchronized 150 ms of audible speech.”

• cites [112] sourcePartly supported

  The Materials and Methods sections on Audio Signal Processing and Decoding Approach describe aligned neural features and 150 ms audio units selected from training data. But ECoG Signal Processing specifies 450 ms neural windows, not 150 ms neural segments. Page numbers are unavailable in our copy. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Multiple studies have demonstrated that isolated phonemes can be decoded from neural activity.

“Multiple studies have demonstrated that isolated phonemes can be decoded from neural activity17,106,107,118 (Fig. 1; Mugler et al. 2014 and Herff et al. 2015).”

• cites [17] sourcePartly supported

  The ‘Phonetic Organization of Spatial Patterns’ section reports distinct neural activity patterns for consonant articulator groups and for three vowels. However, the Methods describe participants producing consonant–vowel syllables, and the reported analysis clusters neural patterns rather than demonstrating decoding of isolated phonemes. No page number is available in our copy. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [106] sourceSupported
• cites [107] sourcePartly supported

  The Introduction reports that earlier studies classified phonemes isolated from words, supporting the general claim. But the Discussion says Mugler et al.’s own study classified individual phonemes within word production, not isolated phonemes, so it does not support citing this study as an example of that specific result. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [118] sourceSupported

Herff et al. developed a concatenative synthesizer that transformed SEEG neural activity into audible speech.

“Herff et al.112 developed a concatenative synthesizer to transform neural activity recorded from SEEG into audible speech (Fig. 1; Herff et al. 2019).”

• cites [112] sourcePartly supported

  The Decoding Approach and Results sections describe a unit-selection method that concatenated speech units to generate audible speech from neural activity. But Materials and Methods (Data Recording) says the activity was recorded with an ECoG electrode grid, not SEEG; the recording modality in the claim is incorrect. The indexed passages provide no page numbers. (LLM, conf=0.92)

  Experimental — accuracy not yet measured

Voice-conversion models require as little as 3 s of recorded speech.

“Voice-conversion models require as little as 3 s of recorded speech126; however, larger personalized training datasets can be leveraged if the individual has additional pre-injury audio recordings.”

• cites [126] sourcePartly supported

  Section 5 (Zero-Shot Voice Conversion; page not provided) reports voice-conversion experiments using target-speaker reference recordings longer than 3 seconds. This supports the use of short recordings, but does not establish that exactly 3 seconds is sufficient or that voice-conversion models generally require only that amount. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Card et al. trained a personalized TTS model on a small corpus of pre-injury speech as an alternative to voice conversion.

“Card et al.16 trained a personalized TTS model on a small corpus of pre-injury speech, which is an alternative to using a voice-conversion model.”

• cites [16] sourcePartly supported

  The Abstract (page unavailable) says the decoded words were vocalized with text-to-speech software designed to sound like the participant’s pre-ALS voice. Our copy does not say that Card et al. trained that TTS model on a small corpus of pre-injury speech or compare it with voice conversion. The short training times reported in the Abstract concern the neural speech decoder, not the TTS voice. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Models that learn an underlying manifold structure for a behaviour can be used to stabilize MEA activity over time.

“One line of work involves developing models that learn an underlying manifold (low-dimensional spaces that capture neural population dynamics) structure for a given behaviour103,104,121, which can then be used to stabilize MEA activity over time.”

• cites [103] sourceSupported
• cites [104] sourcePartly supported

  In “Leveraging manifolds and dynamics to stabilize iBCI decoding” (page unavailable), the paper describes learning a neural manifold and updating the mapping from recorded activity to that manifold across sessions. It supports stabilization of decoding despite changes in electrode recordings, but does not say the method stabilizes MEA activity itself. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [121] sourcePartly supported

  In “Using latent factors and dynamics to increase BMI performance” (page unavailable), the paper describes models of low-dimensional neural activity that help maintain BMI decoding performance across recording sessions and neuron loss. It does not say these models stabilize the MEA-recorded activity itself; rather, they help decoding remain robust as recordings change. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

A text-decoding-specific approach uses large language models to automatically recalibrate MEA-based text neuroprostheses.

“A second approach, specific to text decoding, involves leveraging large language models to automatically recalibrate MEA-based text neuroprostheses.”

• cites [9] sourcePartly supported

  The Daily decoder retraining section reports an offline test in which a language model corrected errors and enabled automatic, unsupervised recalibration of a handwriting decoder. It does not identify that model as a large language model, and it says the approach still needs online confirmation. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  The Methods section (‘Decoding speech’; page unavailable) describes an intracortical brain-to-text neuroprosthesis that uses language models to turn predicted phonemes into sentences. It separately says data from different days were combined to calibrate the decoder. The Discussion (page unavailable) says recalibration used prompted sentences with ground-truth labels; the paper does not support the claim that large language models automatically performed recalibration. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [105] sourceSupported

Metzger et al. first used voice conversion to personalize synthesized speech for an individual with vocal-tract paralysis using pre-injury speech samples.

“Metzger et al.14 first used this voice-conversion approach to personalize synthesized speech for an individual with vocal-tract paralysis using speech samples recorded before the injury.”

• cites [14] sourceSupported

Self-perception of speech is an important component of speech-motor control.

“Furthermore, self-perception of speech is an important component to speech-motor control122,123.”

• cites [122] sourceSupported
• cites [123] sourceSupported

A language model combined neural-based probabilities of each word with linguistic likelihoods of word sequences, improving WERs of the decoded sentences.

“A language model combined the neural-based probabilities of each word with linguistic likelihoods of word sequences, improving WERs of the decoded sentences12 (Fig. 1; Moses et al. 2021; Supplementary Table 1).”

• cites [12] sourceSupported

Using CTC loss, recent work has achieved large-vocabulary decoding while keeping decoding speeds closer to natural speech.

“Recent works have achieved large-vocabulary decoding while maintaining decoding speeds closer to natural speech, by using CTC loss.”

• cites Metzger et al. 2023 sourcePartly supported

  The Overview and Methods > Text decoding describe training the decoders with CTC loss, and Text decoding reports real-time decoding with a 1,024-word vocabulary. However, the 39,378-word result was evaluated offline, while the reported rates of 89.9 and 101 words per minute were for smaller, predefined phrase sets. The passages do not establish that the largest-vocabulary decoding maintained speeds close to natural speech. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites Willett et al. 2023 sourceSupported
• cites Card et al. 2023 sourcePartly supported

  The Results (Online decoding performance; page unavailable) report decoding with a vocabulary of over 125,000 words at the participant’s self-paced rate of 31.6 words per minute. The Methods (Decoding speech; page unavailable) describe phoneme prediction and language models, but the searched passages do not establish use of CTC loss. The Discussion calls the participant’s speaking rate slow, so the paper does not substantiate the comparison to natural speech speed. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

By using CTC loss, Sun et al. trained an RNN to decode character sequences from brain activity during sentence production.

“By using the connectionist temporal classification (CTC) loss119, which scores the loss between predictions from an input sequence (such as neural activity) and an output sequence (such as linguistic subunits) without requiring their alignment, Sun et al.109 trained an RNN to decode character sequences from brain activity during sentence production.”

• cites [119] sourceNot checked — no full text of the source
• cites [109] sourcePartly supported

  The Methods (Brain2Char architecture; page not provided) describe Bi-LSTM recurrent layers in the neural-feature encoder and CTC loss for decoding character sequences from brain recordings. But the text decoder itself uses dilated CNN layers, so describing the trained decoder simply as an RNN misstates that part of the architecture. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

MEA-based decoders, including text decoders, have traditionally needed day-to-day recalibration because of signal instabilities and drift.

“MEA-based decoders, including text decoders9,15, have traditionally required day-to-day recalibration owing to signal instabilities and drift, requiring additional dedicated time from the user.”

• cites [9] sourceSupported
• cites [15] sourceSupported

Anumanchipalli et al.'s approach further improved speech-synthesis performance by leveraging the underlying articulatory organization of the SMC.

“Their approach further improved speech-synthesis performance by leveraging the underlying articulatory organization of the SMC.”

• cites [110] sourceSupported

The manifold-stabilization approach has mainly been applied in upper limb decoders.

“This approach has primarily been applied in the context of upper limb decoders103 but may scale to speech neuroprostheses in the future.”

• cites [103] sourceSupported

Lexicon size may be increased without substantially limiting performance.

“the latter of which may be increased without substantially limiting performance14,15.”

• cites [14] sourceSupported
  Experimental — accuracy not yet measured
• cites [15] sourcePartly supported

  In “Design considerations for speech BCIs” (page unavailable), word error rates levelled off at about 1,000 words, supporting the idea that further vocabulary growth may have little effect. But the paper also says only very small vocabularies retained a large accuracy advantage: in “Decoding attempted speech” (page unavailable), the error rate rose from 9.1% with 50 words to 23.8% with 125,000 words. The unqualified claim understates that loss of performance. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Previous work showed that first decoding articulatory features and transforming them to acoustic features improved synthesized speech quality compared with decoding acoustics directly from neural data.

“Previous work has demonstrated that first decoding articulatory features and transforming them to acoustic features improved the quality of synthesized speech (compared with decoding acoustics directly from neural data)110.”

• cites [110] sourceSupported
  Experimental — accuracy not yet measured

Proxy waveforms generated from TTS or recorded by able speakers can be used as references in MCD computations.

“To resolve this, proxy waveforms, such as waveforms generated from TTS or recorded by able speakers and used during model training, can be used as references in MCD computations14.”

• cites [14] sourceSupported

ASR systems are typically trained to recognize natural waveforms, not artificially synthesized ones.

“ASR can also be used to transcribe synthesized speech into text to compute WERs; however, ASR systems are typically trained to recognize natural waveforms that are not artificially synthesized153.”

• cites [153] sourcePartly supported

  The Methods (Automated speech recognition models; page unavailable) say that the Wav2vec 2.0 model was trained on speech corpora, and the paper applies ASR to BCI-synthesized speech. Our copy does not establish that ASR systems generally are trained on natural waveforms rather than artificially synthesized ones. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

In previous multiparticipant studies in able speakers, participants had near-identical speech capacity.

“By contrast, in previous multiparticipant studies in able speakers, participants had near-identical speech capacity106,108,136–138.”

• cites [106] sourcePartly supported

  The Subjects and Experiment sections (2.1 and 2.3; page unavailable) describe seven participants who produced speech by reading text aloud, supporting the characterization of this as a multiparticipant study of speakers. They do not assess or establish that the participants had near-identical speech capacity. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [108] sourcePartly supported

  The Methods (Task; page not provided) say that participants read sentences aloud, supporting the description of this as a multiparticipant study of able speakers. Neither that section nor the Methods participant description (page not provided) establishes that their speech capacities were near-identical. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [136] sourceNot checked — no full text of the source
• cites [137] sourcePartly supported

  The Results describe seven participants completing overt word-production tasks, and report that they could produce both real words and pseudowords. Neither the Results nor Materials and Methods compare participants’ speech capacities or establish that they were near-identical; a search for speech ability and deficits found no such comparison. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [138] sourcePartly supported

  The Online Methods (Participants; page unavailable) describe 16 participants selected for cognitive and language abilities in the average range or above, supporting the description of a multiparticipant study of able speakers. That criterion does not establish that their speech capacities were near-identical. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Studies with ECoG and MEAs in individuals with paralysis demonstrated feasibility of decoding articulatory features.

“Studies with both ECoG14 and MEAs15,16,52 in individuals with paralysis have demonstrated the feasibility of decoding articulatory features”

• cites [14] sourceSupported
• cites [15] sourceSupported
• cites [16] sourcePartly supported

  The Methods (Decoding speech; page unavailable) describe using cortical neural activity to predict attempted phonemes, and the Methods (Recording array implant locations and decoding contributions; page unavailable) identify four microelectrode arrays implanted in speech-related cortex. The study supports the feasibility of MEA-based speech decoding in a person with paralysis, but it reports decoding phonemes rather than articulatory features specifically. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [52] sourcePartly supported

  The Results (‘Speech can be decoded from intracortical activity on individual trials’; page unavailable) report decoding spoken syllables and words from intracortical array recordings in two people with tetraplegia. The paper relates the neural patterns to articulation, but it does not report decoding articulatory features themselves; its Discussion says linking activity to continuous articulatory movements requires future work. This source supports the MEA portion of the sentence, not an independent ECoG demonstration. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

TTS system performance is often described using mel-cepstral distortion and open-ended human-transcribed WER.

“The performance of TTS systems is often described using mel-cepstral distortion (MCD) and open-ended human-transcribed WER149,150.”

• cites [149] sourceCould not verify

  The Approach and Results sections evaluate AAC word prediction using communication rate, keystrokes, and related measures. Searches for mel-cepstral distortion, TTS, and human-transcribed word error rate found no passage addressing how TTS performance is described. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [150] sourceNot checked — no full text of the source

Delayed auditory feedback disrupts speech production.

“Delayed auditory feedback disrupts speech production, so an ideal streaming neural-speech synthesizer would output sound with negligible latencies similar to the self-perception of able speech (within a few milliseconds).”

• cites [154] sourceCould not verify

  Section 3.8 (p. 9) describes Whisper’s 30-second input chunks and buffered transcription, not the effects of delayed auditory feedback on speech production. Searches for delayed auditory feedback and millisecond-scale synthesis latency found no passage in our copy addressing the attributed claim. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Metzger et al. demonstrated that decoded articulatory features can animate a digital avatar face personalized to resemble the individual.

“Metzger et al. demonstrated that it is possible to use these decoded articulatory features to animate a digital avatar face, which was personalized to resemble the likeness of the individual14 (Fig. 3d).”

• cites [14] sourcePartly supported

  The Facial-avatar decoding section says decoded articulatory gestures were used to animate the avatar face, and the Discussion calls it a personalized avatar. However, our copy does not say the avatar’s face was personalized to resemble the participant; the explicit resemblance claim in Methods > Speech synthesis concerns her synthetic voice. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

AAC typically do not exceed 15 WPM.

“This would likely provide substantial improvements over AAC, which typically do not exceed 15 WPM (refs. 146–148).”

• cites [146] sourceSupported
• cites [147]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [148] sourceNot checked — no full text of the source

Generative pretrained transformers are a class of language models particularly powerful for rescoring decoded sequences.

“Generative pretrained transformers145 are a class of language models that are particularly powerful for this application.”

• cites [145] sourcePartly supported

  The Abstract (p. 1) identifies GPT-4 as a Transformer-based model pre-trained to predict the next token, supporting the description of GPTs as language models. The report does not establish that they are particularly powerful for rescoring decoded sequences; searches for rescoring and speech-recognition decoding found no supporting passage. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Most early studies in individuals with vocal-tract paralysis involve only single participants.

“Given that most early studies in individuals with vocal-tract paralysis involve only single participants12–16,64,95,120, great care is needed when comparing results, as the broad spectrum of disease presentations is not sampled”

• cites [12] sourceSupported
• cites [13] sourceSupported
• cites [14] sourceSupported
• cites [15] sourceSupported
• cites [16] sourceSupported
• cites [64] sourcePartly supported

  The Introduction and Experimental model and subject details describe a study of one participant. But the Discussion calls him “speech-abled,” so this paper does not provide an example of a study in an individual with vocal-tract paralysis. It also does not establish that most early studies used single participants. (LLM, conf=0.90)

  Experimental — accuracy not yet measured
• cites [95] sourcePartly supported

  The Experimental Section (Clinical Trial and Participant; page unavailable) reports data from one participant, so this paper supports the single-participant aspect of the claim. It describes severe dysarthria from ALS but says the participant could still phonate and articulate; it does not establish vocal-tract paralysis or the broader assertion that most early studies used single participants. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [120] sourceSupported

Articulatory features may be useful as a standalone output space for speech and non-speech orofacial gestures.

“In addition, articulatory features may be useful as a standalone output space for both speech and non-speech orofacial gestures14,52,113,114 (Fig. 3d).”

• cites [14] sourceSupported
• cites [52] sourcePartly supported

  The Results (Speech-related activity in dorsal motor cortex; page unavailable) show that neural activity grouped phonemes by place of articulation and that many electrodes active during speech were also active during non-speech orofacial movements. This supports the relevance of articulatory features to both behaviors, but the paper does not evaluate them as a standalone output space; the Methods say precise movement kinematics were not collected. (LLM, conf=0.81)

  Experimental — accuracy not yet measured
• cites [113] sourcePartly supported

  The Discussion (page unavailable) reports that distinct articulator and tongue movements can be classified and suggests their neural activity could provide additional BCI control signals. It also says the findings may benefit speech-BCI development, but the study tested isolated movements rather than a standalone articulatory output space for both speech and non-speech gestures. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [114] sourcePartly supported

  The Discussion (page not provided) reports that neural activity can distinguish posed facial expressions involving the lips, jaw, nose, and eyebrows. The Conclusion proposes adding those expressions to communication BCIs, supporting a potential output for non-speech gestures. The paper does not establish articulatory features as a standalone output space for speech. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Speech-detection algorithms are highly successful in correctly identifying volitional attempts of users to overtly and silently speak.

“an additional safeguard is to use a speech-detection algorithm to gate any decoding process, which is highly successful in correctly identifying volitional attempts of users to overtly and silently speak12,120.”

• cites [12] sourcePartly supported

  The Results, ‘Word Detection and Classification’ (page not provided), report that the detector identified 98% of one participant’s attempted word productions, supporting high detection of attempted speech. The study involved a person with anarthria; it does not establish comparable performance for both overt and silent speech across users. It also reports 968 spurious detections, so the result does not establish that detected events consistently represent volitional attempts. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [120] sourcePartly supported

  The Speech detection section (page unavailable) describes a real-time model that detects the participant’s silent-speech attempts to engage the spelling system, supporting the gating part of the claim. It does not establish that detection was highly accurate for both overt and silent attempts; the overt-versus-silent results concern classification of attempted code words, not detection of whether speech was attempted. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Higher-density cortical sampling via increased surface or penetrating electrodes has been shown to increase decoding model performance.

“For example, higher-density cortical sampling — via an increased number of surface or penetrating electrodes — has been shown to increase the performance of decoding models14–16,175,176.”

• cites [14] sourceSupported
• cites [15] sourcePartly supported

  In “Design considerations for speech BCIs” (Fig. 4b; page unavailable), the paper reports that offline decoding accuracy improved as more intracortical electrodes were included. It suggests denser arrays may improve future accuracy, but does not demonstrate the claim for surface electrodes or test a newly increased physical electrode density. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  The Methods say this study doubled the number of implanted microelectrodes relative to an earlier study, and the Discussion reports higher decoding performance. But the Discussion calls electrode count only a possible contributing factor and names other changes that may have helped; it does not establish that more electrodes caused the improvement. This paper also does not test increased surface-electrode sampling. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [175] sourceNot checked — no full text of the source
• cites [176] sourceSupported

A spelling framework using attempted speech or handwriting can facilitate a similar tradeoff, allowing users to spell arbitrary sentences with high accuracy but lower speeds.

“A similar tradeoff can be facilitated by having a spelling framework available on the neuroprosthesis, either using attempted speech120 or handwriting9, which would allow users to spell out arbitrary sentences with high accuracy (albeit with substantially lower speeds).”

• cites [120] sourceSupported
• cites [9] sourcePartly supported

  The Discussion (page unavailable) says the handwriting BCI lets its user express any sentence, and Decoding handwritten sentences reports high accuracy. The paper does not establish the claimed tradeoff of substantially lower speed; it describes 90 characters per minute as a new BCI speed benchmark. The attempted-speech alternative has a separate citation and is not assessed here. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Long and recurring training periods for speech neuroprostheses might discourage long-term use.

“As long and recurring training periods for speech neuroprostheses might discourage long-term use158,159, it is essential to develop decoders that are stable over time or can be quickly recalibrated102,104,105 to avoid dedicated and high-effort re-training periods inconveniencing users.”

• cites [158] sourceNot checked — no full text of the source
• cites [159] sourcePartly supported

  The Results (page not available) report users’ concerns that BCI setup was time-consuming and training-intensive for caregivers, particularly with frequent staff turnover. The paper does not say that speech-neuroprosthesis users face long, recurring training periods or that such periods discourage long-term use. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

During online inference, decoded discrete acoustic-speech unit sequences were vocoded into sentence-level speech intelligible to untrained human listeners.

“During online inference, the decoded discrete acoustic-speech unit sequences were vocoded into sentence-level speech that was intelligible to untrained human listeners (Supplementary Table 1).”

• cites Supplementary Table 1Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Persisting neural representations in individuals with dysarthria or anarthria provide a substrate for decoding intended articulation and vocalization even if motor commands no longer reach the vocal tract.

“These persisting neural representations provide a substrate for decoding intended articulation and vocalization, even if these motor commands no longer reach their destination in the vocal tract.”

• AssumptionRests on an assumption: In individuals with dysarthria or anarthria, cerebral cortex neuronal populations that contain representations of articulatory movements may remain intact.

Speech neuroprostheses already achieve performance levels that exceed the communication rates of current assistive-communication technology.

“Although restoring natural speech is a long-term goal, speech neuroprostheses already have performance levels that surpass communication rates offered by current assistive-communication technology.”

All invasive recording techniques share risks related to anaesthesia, craniotomy, dural opening, and the possibility of infection.

“However, all invasive recording techniques share risks related to the anaesthesia, craniotomy and dural opening in the implantation surgery and the possibility of infection81.”

• cites [81]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

ECoG has been used clinically for decades for localizing seizures and brain mapping in surgical work-up for epilepsy treatment.

“ECoG has been used clinically for decades for localizing seizures and for brain mapping in surgical work-up for epilepsy treatment90.”

• cites [90] sourceNot checked — no full text of the source

Stable acquisition, broad coverage and high performance of ECoG make it a strong candidate for a clinically viable neuroprosthesis.

“Thus, the stable acquisition, broad coverage and high performance of ECoG make it a strong candidate for a clinically viable neuroprosthesis.”

Together, these studies highlighted a large region of cortex, primarily on the SMC, that has been subsequently targeted in the first clinical trials of speech neuroprostheses.

“Together, these studies all highlighted a large region of cortex, primarily on the SMC, that has been subsequently targeted in the first clinical trials of speech neuroprostheses.”

Speech and language features encoded in the STG, SMG and Broca's area could potentially be leveraged for speech decoding in disorders where articulatory representations on the SMC are not intact.

“Thus, speech and language features encoded in the STG, SMG and Broca’s area could potentially be leveraged for speech decoding in disorders in which articulatory representations on the SMC are not intact (Box 1).”

Previously described speech-synthesis approaches required precise alignment between neural activity and ground-truth acoustics.

“The previously described speech-synthesis approaches required precise alignment between neural activity and ground-truth acoustics.”

These approaches are all limited by restrictive, predefined vocabulary sizes.

“Despite their successes, these approaches are all limited by restrictive, predefined vocabulary sizes.”

A key benefit of CTC loss is that the user can produce a phrase more naturally without linking isolated speech attempts to individual words.

“A key advantage of using CTC loss is that the individual can more naturally produce a phrase in this approach, without having to associate isolated speech attempts with individual words.”

CTC-based approaches have led to state-of-the-art decoding performance for individuals with paralysis.

“Overall, this has led to state-of-the-art decoding performance for individuals with paralysis.”

Because speech is variable across people and expressive, speech-synthesis neuroprostheses are well suited for personalization.

“Because speech is variable across people and expressive, neuroprostheses that synthesize speech are well suited for personalization (Fig. 3c).”

Language models converted decoded phoneme sequences into words and sentences in these studies.

“Language models then mapped decoded phoneme sequences into words and sentences (Fig. 3b and Supplementary Table 1).”

• cites Fig. 3bNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Supplementary Table 1Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

A voice-conversion model can convert decoded speech waveforms into personalized waveforms resembling the user.

“A voice-conversion model can be applied to convert decoded speech waveforms into personalized waveforms that resemble the likeness of the user.”

Decoders can generalize to larger vocabularies by targeting low-dimensional subword units such as phonemes.

“Furthermore, decoders can generalize to larger vocabularies by targeting low-dimensional subword units (such as phonemes) that can be built into a much higher-dimensional space of words and sentences.”

Promising avenues exist to create MEA-based text neuroprostheses that do not require daily recalibration time from the user.

“Thus, promising avenues exist to create MEA-based text neuroprostheses that do not require daily recalibration time from the user.”

Individuals with severe dysarthria or anarthria may lack the ability to make any intelligible vocalizations, especially for longer sentences.

“However, individuals with severe cases of dysarthria or anarthria may lack the ability to make any intelligible vocalizations, especially for longer sentences.”

CTC loss can circumvent the need for alignment by training models to map neural activity sequences to acoustic signal sequences that can be vocoded into speech.

“Similar to text decoding, a way to circumvent the need for alignment is to use CTC loss to train models that map input sequences of neural activity to output sequences of acoustic signals that may then be vocoded into speech.”

A major challenge for speech neuroprostheses in vocal-tract paralysis is the lack of intelligible ground-truth acoustic signals to align with neural data during training.

“However, as described earlier, a major challenge of creating speech neuroprostheses for individuals with vocal-tract paralysis is the lack of intelligible ground-truth acoustic signals to align with the neural data during training.”

The pseudo-label approach allows automatic RNN recalibration without additional training labels or dedicated user time.

“This allows the RNN to be automatically recalibrated without additional training labels or dedicated time on the part of an individual with paralysis.”

For individuals who cannot speak, there may be no way to reliably align neural activity with ground-truth produced acoustics.

“However, for individuals who cannot speak, there may be no way to reliably align neural activity with ground-truth produced acoustics.”

Given the strong link between speech acoustics and underlying vocal-tract movements, articulation is a relevant feature space for speech decoding and synthesis.

“Given the strong link between speech acoustics and underlying vocal-tract movements, articulation constitutes another relevant feature space (Fig. 3d) for speech decoding and speech synthesis128,129.”

• cites [128] sourceNot checked — no full text of the source
• cites [129] sourceNot checked — no full text of the source

When attempting to speak, audible vocalizations produced by an individual with anarthria are generally unintelligible because of lack of coordinated vocal-tract control.

“When attempting to speak, any audible vocalizations that an individual with anarthria produces are generally unintelligible140 owing to their lack of coordinated vocal-tract control (Fig. 4a).”

• cites [140]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

High performance on an evaluation set reflective of natural language containing words unseen during training is a strong indicator that a speech neuroprosthesis would allow the user to produce words beyond the training vocabulary.

“However, a strong indicator that a speech neuroprosthesis would allow the user to produce words beyond the training vocabulary is high performance on an evaluation set that is reflective of natural language, containing words not seen by the speech-decoding model during training.”

MCD is preferred to correlation-based metrics because it weighs features most perceptually relevant for conveying information.

“It is preferred to correlation-based metrics, such as computing a Pearson correlation between ground truth and decoded acoustic waveforms, as MCD weighs the features in the produced sound that are most perceptually relevant for conveying information151.”

• cites [151] sourceNot checked — no full text of the source

Language models can rescore sequences of linguistic features decoded from neural activity by evaluating how likely decoded sequences are in natural language.

“Most prominently, language models can be used to rescore sequences of linguistic features (such as phonemes or words) that are decoded from neural activity. Language models perform this re-scoring by evaluating how likely decoded sequences are to occur in natural language.”

MCD use is complicated by lack of precisely aligned ground-truth speech for individuals who cannot speak.

“The use of the MCD metric is, however, complicated by the lack of precisely aligned ground-truth speech when working with individuals who cannot speak.”

Current speech neuroprostheses do not decode or target internal monologue.

“Although current speech neuroprostheses do not decode or target internal monologue (Fig. 4a), an additional safeguard is to use a speech-detection algorithm to gate any decoding process, which is highly successful in correctly identifying volitional attempts of users to overtly and silently speak12,120.”

• cites Fig. 4aNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

A positive correlation exists between decoder performance and electrode density and/or relevant coverage afforded by the neural interface.

“Overall, a positive correlation exists between decoder performance and electrode density and/or the relevant coverage afforded by the corresponding neural interface, emphasizing that improvements to these two hardware aspects can lead to improved speech brain computer interfaces, regardless of the specific recording modality.”

Improvements to electrode density and relevant coverage can lead to improved speech brain-computer interfaces regardless of recording modality.

“Overall, a positive correlation exists between decoder performance and electrode density and/or the relevant coverage afforded by the corresponding neural interface, emphasizing that improvements to these two hardware aspects can lead to improved speech brain computer interfaces, regardless of the specific recording modality.”