{"claims": [{"text": "9.1% word error rate on a 50-word vocabulary", "quote_or_locator": "Abstract and Figure 2b, Table 1: 'T12 achieved a 9.1% word error rate for the 50-word vocabulary across all vocalizing days'"}, {"text": "2.7 times fewer errors than the previous state-of-the-art speech BCI", "quote_or_locator": "Abstract: '2.7 times fewer errors than the previous state-of-the-art speech BCI'"}, {"text": "23.8% word error rate on a 125,000-word vocabulary", "quote_or_locator": "Abstract and Figure 2b, Table 1: 'a 23.8% word error rate on a 125,000-word vocabulary'"}, {"text": "First successful demonstration of large-vocabulary decoding", "quote_or_locator": "Abstract and main text: 'the first successful demonstration, to our knowledge, of large-vocabulary decoding'"}, {"text": "62 words per minute decoding speed", "quote_or_locator": "Abstract and Figure 2c: 'T12's attempted speech was decoded at 62 words per minute'"}, {"text": "3.4 times faster than the previous record", "quote_or_locator": "Abstract: 'which is 3.4 times as fast as the previous record'"}, {"text": "Previous state-of-the-art speed was 18 words per minute for a handwriting BCI", "quote_or_locator": "Main text: 'more than triples the speed of the previous state of the art for any type of BCI (18 words per minute for a handwriting BCI)'"}, {"text": "Area 6v achieved 92% accuracy decoding 33 orofacial movements", "quote_or_locator": "Main text and Figure 1d: 'we could decode from among 33 orofacial movements with 92% accuracy'"}, {"text": "Area 6v achieved 62% accuracy decoding 39 phonemes", "quote_or_locator": "Main text and Figure 1d: '39 phonemes with 62% accuracy'"}, {"text": "Area 6v achieved 94% accuracy decoding 50 words", "quote_or_locator": "Main text and Figure 1d: '50 words with 94% accuracy'"}, {"text": "Area 44 showed classification accuracy below 12% for movements, phonemes, and words", "quote_or_locator": "Main text: 'area 44 appeared to contain little to no information about orofacial movements, phonemes or words (classification accuracy below 12%)'"}, {"text": "Tuning to speech articulators was intermixed at single-electrode level", "quote_or_locator": "Main text: 'tuning to speech articulators (jaw, larynx, lips or tongue) was intermixed at the single-electrode level'"}, {"text": "All speech articulators represented within both 3.2 × 3.2 mm² arrays", "quote_or_locator": "Main text: 'all speech articulators were clearly represented within both 3.2 × 3.2 mm² arrays'"}, {"text": "Correlation between electromagnetic articulography and neural data for consonants was 0.61", "quote_or_locator": "Main text and Figure 3b: 'the correlation between electromagnetic articulography (EMA) and neural data was 0.61, far above chance'"}, {"text": "Phoneme error rate was 19.7% for vocal speech", "quote_or_locator": "Main text and Table 1: 'Phoneme error rates computed on the raw RNN output were 19.7% for vocal speech'"}, {"text": "Training dataset consisted of 10,850 total sentences by final day", "quote_or_locator": "Main text: 'By the final day our training dataset consisted of 10,850 total sentences'"}, {"text": "Data collection and RNN training lasted 140 minutes per day on average", "quote_or_locator": "Main text: 'Data collection and RNN training lasted for 140 min per day on average (including breaks)'"}, {"text": "Offline analysis with improved language model reduced error rate to 17.4%", "quote_or_locator": "Main text and Table 1: 'an improved language model could decrease word error rates from 23.8 to 17.4%'"}, {"text": "Offline analysis with improved language model and proximal test set reduced error rate to 11.8%", "quote_or_locator": "Main text and Table 1: 'testing on more proximal sentences further decreased word error rates to 11.8%'"}, {"text": "Performance was good without using training data on the new day, with 30% word error rate", "quote_or_locator": "Main text and Figure 4c: 'word error rate was 30% with no retraining'"}, {"text": "Doubling electrode count cuts error rate nearly in half", "quote_or_locator": "Main text and Figure 4b: 'doubling the electrode account appears to cut the error rate nearly in half'"}, {"text": "Participant T12 has bulbar-onset ALS and retains limited orofacial movement", "quote_or_locator": "Main text: 'Our participant (T12) has bulbar-onset amyotrophic lateral sclerosis (ALS) and retains some limited orofacial movement and an ability to vocalize, but is unable to produce intelligible speech'"}, {"text": "Arrays were placed in area 6v (ventral premotor cortex) and area 44 (part of Broca's area)", "quote_or_locator": "Main text: 'we recorded neural activity from four microelectrode arrays—two in area 6v (ventral premotor cortex) and two in area 44 (part of Broca's area)'"}, {"text": "Silent speech achieved 11.2% word error rate on 50-word vocabulary", "quote_or_locator": "Main text and Table 1: 'T12 achieved a 9.1% word error rate for the 50-word vocabulary across all vocalizing days (11.2% for silent)'"}, {"text": "Silent speech achieved 24.7% word error rate on 125,000-word vocabulary", "quote_or_locator": "Main text and Table 1: 'a 23.8% word error rate for the 125,000-word vocabulary across all vocalizing days (24.7% for silent)'"}, {"text": "Ventral array contributed more to decoding than dorsal array", "quote_or_locator": "Main text: 'We found that speech could be more accurately decoded from the ventral array, especially during the instructed delay period'"}, {"text": "Reduction in word error rate from 32% to 21% when adding dorsal to ventral array", "quote_or_locator": "Main text: 'offline analyses showed a reduction in word error rate from 32 to 21% when adding the dorsal to the ventral array'"}], "prompt_version": "p1.0", "verdicts": [{"claim": "9.1% word error rate on a 50-word vocabulary", "verdict": "supported", "evidence": "T12 achieved a 9.1% word error rate for the 50-word vocabulary across all vocalizing days", "note": null}, {"claim": "2.7 times fewer errors than the previous state-of-the-art speech BCI", "verdict": "supported", "evidence": "2.7 times fewer errors than the previous state-of-the-art speech BCI", "note": null}, {"claim": "23.8% word error rate on a 125,000-word vocabulary", "verdict": "supported", "evidence": "a 23.8% word error rate on a 125,000-word vocabulary", "note": null}, {"claim": "First successful demonstration of large-vocabulary decoding", "verdict": "supported", "evidence": "the first successful demonstration, to our knowledge, of large-vocabulary decoding", "note": null}, {"claim": "62 words per minute decoding speed", "verdict": "supported", "evidence": "T12 spoke at an average pace of 62 words per minute", "note": null}, {"claim": "3.4 times faster than the previous record", "verdict": "supported", "evidence": "which is 3.4 times as fast as the previous record", "note": null}, {"claim": "Previous state-of-the-art speed was 18 words per minute for a handwriting BCI", "verdict": "supported", "evidence": "more than triples the speed of the previous state of the art for any type of BCI (18 words per minute for a handwriting BCI)", "note": null}, {"claim": "Area 6v achieved 92% accuracy decoding 33 orofacial movements", "verdict": "supported", "evidence": "we could decode from among 33 orofacial movements with 92% accuracy", "note": null}, {"claim": "Area 6v achieved 62% accuracy decoding 39 phonemes", "verdict": "supported", "evidence": "39 phonemes with 62% accuracy", "note": null}, {"claim": "Area 6v achieved 94% accuracy decoding 50 words", "verdict": "supported", "evidence": "50 words with 94% accuracy", "note": null}, {"claim": "Area 44 showed classification accuracy below 12% for movements, phonemes, and words", "verdict": "supported", "evidence": "area 44 appeared to contain little to no information about orofacial movements, phonemes or words (classification accuracy below 12%)", "note": null}, {"claim": "Tuning to speech articulators was intermixed at single-electrode level", "verdict": "supported", "evidence": "tuning to speech articulators (jaw, larynx, lips or tongue) was intermixed at the single-electrode level", "note": null}, {"claim": "All speech articulators represented within both 3.2 × 3.2 mm² arrays", "verdict": "supported", "evidence": "all speech articulators were clearly represented within both 3.2 × 3.2 mm² arrays", "note": null}, {"claim": "Correlation between electromagnetic articulography and neural data for consonants was 0.61", "verdict": "supported", "evidence": "the correlation between electromagnetic articulography (EMA) and neural data was 0.61, far above chance", "note": null}, {"claim": "Phoneme error rate was 19.7% for vocal speech", "verdict": "supported", "evidence": "Phoneme error rates computed on the raw RNN output were 19.7% for vocal speech", "note": null}, {"claim": "Training dataset consisted of 10,850 total sentences by final day", "verdict": "supported", "evidence": "By the final day our training dataset consisted of 10,850 total sentences", "note": null}, {"claim": "Data collection and RNN training lasted 140 minutes per day on average", "verdict": "supported", "evidence": "Data collection and RNN training lasted for 140 min per day on average (including breaks)", "note": null}, {"claim": "Offline analysis with improved language model reduced error rate to 17.4%", "verdict": "supported", "evidence": "an improved language model could decrease word error rates from 23.8 to 17.4%", "note": null}, {"claim": "Offline analysis with improved language model and proximal test set reduced error rate to 11.8%", "verdict": "supported", "evidence": "testing on more proximal sentences further decreased word error rates to 11.8%", "note": null}, {"claim": "Performance was good without using training data on the new day, with 30% word error rate", "verdict": "supported", "evidence": "word error rate was 30% with no retraining", "note": null}, {"claim": "Doubling electrode count cuts error rate nearly in half", "verdict": "supported", "evidence": "doubling the electrode account appears to cut the error rate nearly in half", "note": null}, {"claim": "Participant T12 has bulbar-onset ALS and retains limited orofacial movement", "verdict": "supported", "evidence": "Our participant (T12) has bulbar-onset amyotrophic lateral sclerosis (ALS) and retains some limited orofacial movement and an ability to vocalize, but is unable to produce intelligible speech", "note": null}, {"claim": "Arrays were placed in area 6v (ventral premotor cortex) and area 44 (part of Broca's area)", "verdict": "supported", "evidence": "we recorded neural activity from four microelectrode arrays—two in area 6v (ventral premotor cortex) and two in area 44 (part of Broca's area)", "note": null}, {"claim": "Silent speech achieved 11.2% word error rate on 50-word vocabulary", "verdict": "supported", "evidence": "T12 achieved a 9.1% word error rate for the 50-word vocabulary across all vocalizing days (11.2% for silent)", "note": null}, {"claim": "Silent speech achieved 24.7% word error rate on 125,000-word vocabulary", "verdict": "supported", "evidence": "a 23.8% word error rate for the 125,000-word vocabulary across all vocalizing days (24.7% for silent)", "note": null}, {"claim": "Ventral array contributed more to decoding than dorsal array", "verdict": "supported", "evidence": "We found that speech could be more accurately decoded from the ventral array, especially during the instructed delay period", "note": null}, {"claim": "Reduction in word error rate from 32% to 21% when adding dorsal to ventral array", "verdict": "supported", "evidence": "offline analyses showed a reduction in word error rate from 32 to 21% when adding the dorsal to the ventral array", "note": null}]}