howclose.to
SPEECH BRAIN INTERFACES

The problemsBiology & HealthSpeech BCI

Updated · July 2026

How close are we to giving speech back through a brain implant?Can an implant turn thoughts back into speech?

Can an implant turn thoughts back into speech?How close are we to giving speech back through a brain implant?

Implanted electrodes can already turn neural activity into words on a screen; the open questions are speed, vocabulary, and whether it lasts for years.Electrodes read the brain's signals for speech. The race is to make them fast, accurate, and durable enough for daily life.

Current stageEngineeringEngineering a system that works outside the labBuilding a system that works outside the lab
Read the evidence Where we stand
The path so far

Recorded milestones

  1. Human cortical switch

  2. Independent home communication

Open tests · conceptual branches

TODAY
  1. OPEN TEST02 ↗
    Restore a natural voiceReal-time expressive voiceStreaming and instantaneous systems exist, but direct neural voice still had 43.6% WER in one 2025 evaluation.Synthesize intelligible personalized speech with conversational latency, prosody and unconstrained vocabulary.
  2. OPEN TEST03 ↗
    Work every day at homeIndependent multi-year home useOne participant logged 3,801 hours over 19 months, with care partners still handling connections and startup.Operate at home for years with little calibration or researcher support and stable useful communication.
  3. OPEN TEST04 ↗
    Work for many peopleReproducible clinical benefitCurrent speech records are dominated by single-participant reports and investigational implants.Demonstrate safe, durable benefit across dozens of users and disease anatomies in a regulated multicenter trial.
TODAY
OPEN TEST02 ↗
Restore a natural voiceReal-time expressive voiceStreaming and instantaneous systems exist, but direct neural voice still had 43.6% WER in one 2025 evaluation.Synthesize intelligible personalized speech with conversational latency, prosody and unconstrained vocabulary.
OPEN TEST03 ↗
Work every day at homeIndependent multi-year home useOne participant logged 3,801 hours over 19 months, with care partners still handling connections and startup.Operate at home for years with little calibration or researcher support and stable useful communication.
OPEN TEST04 ↗
Work for many peopleReproducible clinical benefitCurrent speech records are dominated by single-participant reports and investigational implants.Demonstrate safe, durable benefit across dozens of users and disease anatomies in a regulated multicenter trial.
Milestones are in recorded order. Open tests are separate questions, not a forecast.Milestones are in date order. Open tests are separate tests, not a promise.
Speech BCIWhere we stand

The path shows Independent home communication. Open thresholds are separate tests in a conceptual map, not a forecast. No single percentage represents progress.

01 · Where we standWhere we stand

Where we standWhere we stand

The current stage, the evidence recorded, and the single threshold that gates the next stage.The stage reached, the evidence recorded, and the one test still blocking the next step.

What the evidence has shownWhat the evidence has shown

Electrodes read the brain's signals for speech. The race is to make them fast, accurate, and durable enough for daily life.Implanted electrodes can already turn neural activity into words on a screen; the open questions are speed, vocabulary, and whether it lasts for years.

1 of 4 tests recorded1 of 4 thresholds achieved open
✓Say more than fixed phrasesOpen-vocabulary brain-to-textAchieved
Restore a natural voiceReal-time expressive voiceOpen test
Work every day at homeIndependent multi-year home useOpen test
Work for many peopleReproducible clinical benefitOpen test

The next testBlocking threshold

Restore a natural voiceReal-time expressive voice

Achieve under-10% listener WER on unseen spontaneous sentences with independently controlled pitch and emphasis.Achieve under-10% listener WER on unseen spontaneous sentences with independently controlled pitch and emphasis.

02 · Thresholds & evidenceThresholds & evidence

What would count as proof?What would count as proof?

A clear test, the evidence behind it, and the limits of the claim.A clear test, the evidence behind it, and what it still leaves open.

✓Achieved · Aug 2023

Say more than fixed phrasesOpen-vocabulary brain-to-text

Systems now decode from about 125,000 words in individual participants.Decode unconstrained attempted speech from a vocabulary large enough for ordinary English at useful speed.

Recorded result One participant reached 62 words/minute with 23.8% error on a 125,000-word vocabulary.

Next test Replicate below-5% word error across at least ten users during spontaneous conversation.

SOURCE RECORD

2023 large-vocabulary intracortical speech BCI ↗

Still to demonstrateWhat this does not establishA separate system can still have different conditions and results.The named measurement has no wider implication until the supporting system test is also shown.

View every evidence record

Achieved

Say more than fixed phrasesOpen-vocabulary brain-to-text

Systems now decode from about 125,000 words in individual participants.Decode unconstrained attempted speech from a vocabulary large enough for ordinary English at useful speed.

Recorded result: Aug 2023 · One participant reached 62 words/minute with 23.8% error on a 125,000-word vocabulary.

Next test: Replicate below-5% word error across at least ten users during spontaneous conversation.

2023 large-vocabulary intracortical speech BCI ↗

Open test

Restore a natural voiceReal-time expressive voice

Streaming and instantaneous systems exist, but direct neural voice still had 43.6% WER in one 2025 evaluation.Synthesize intelligible personalized speech with conversational latency, prosody and unconstrained vocabulary.

Next test: Achieve under-10% listener WER on unseen spontaneous sentences with independently controlled pitch and emphasis.

Open test

Work every day at homeIndependent multi-year home use

One participant logged 3,801 hours over 19 months, with care partners still handling connections and startup.Operate at home for years with little calibration or researcher support and stable useful communication.

Next test: Repeat two-year, 1,000-hour-plus use in several participants using a fully implanted wireless device.

Open test

Work for many peopleReproducible clinical benefit

Current speech records are dominated by single-participant reports and investigational implants.Demonstrate safe, durable benefit across dozens of users and disease anatomies in a regulated multicenter trial.

Next test: A prospective multicenter cohort must publish user-level accuracy, complications, explants and daily-use outcomes.

03 · The recordThe record

The record behind the verdictThe record behind the verdict

Major milestones and smaller context events share one source-linked record. Scheduled targets are marked separately from recorded events.Milestones, smaller events and setbacks share one sourced record. Scheduled targets are marked separately from recorded events.

Event key
TheoryExperimentFundingSetbackPolicyDeployment Future / scheduled
1973-19971 event1 shown

The Direct-Link Hypothesis

Recorded event: “Direct brain-computer link proposed”.

1973
Direct brain-computer link proposedTheory
Jacques Vidal formalized direct brain-computer communication using measurable electrical brain signals and online computing.
Jacques Vidal and UCLA
1998-20083 events2 shown

Single Channels Move Cursors

This record spans “Locked-in patient controls switch” to “BrainGate proves multidimensional control”.

1998
Locked-in patient controls switchExperiment
A neurotrophic cortical electrode recorded action potentials for months and enabled one woman with ALS to make binary selections.
Philip Kennedy and Roy Bakay
2004
Neural typing reaches three characters
A locked-in participant moved a cursor over a virtual keyboard at only 3 characters per minute using an implanted electrode.
Neural Signals researchers
2006
BrainGate proves multidimensional controlExperiment
A 96-channel array let a person with tetraplegia open email, operate a television and move a robotic hand using imagined movement.
BrainGate consortium
2009-20182 events0 shown

Motor Intent Becomes Language

This record spans “Implant drives vowel synthesizer” to “Cursor typing accelerates”.

2009
Implant drives vowel synthesizer
Two neural channels predicted vowel formant frequencies in real time; the system produced vowel-like sounds, not words.
Frank Guenther and colleagues
2017
Cursor typing accelerates
Three participants reached 6.5, 6.3 and 2.7 correct characters per second with a point-and-click intracortical BCI.
Chethan Pandarinath and colleagues
2019-20223 events2 shown

Neural Decoders Learn Speech

This record spans “Cortical activity synthesizes sentences” to “Handwriting bypass reaches phone speed”.

2021
Anarthric speech decoded onlineExperiment
One participant decoded a 50-word vocabulary at 15.2 words/minute and 25.6% median word error with language modeling.
David Moses and UCSF
2021
Handwriting bypass reaches phone speed
Attempted handwriting produced 90 characters/minute at 94.1% raw online accuracy in one participant.
Francis Willett and BrainGate
2023-203614 events4 shown

Records Confront Daily Life

This record spans “Percutaneous risks stay visible” to “Six-year follow-up scheduled” and includes scheduled targets.

2023
Percutaneous risks stay visible
BrainGate safety data included one localized pedestal-site infection treated with oral antibiotics; devices remained investigational.
BrainGate consortium
2023
Large vocabulary reaches sixty-twoExperiment
One ALS participant decoded 125,000 words at 62 words/minute with 23.8% word error in prompted laboratory speech.
Francis Willett and Stanford BrainGate
2024
Seven-year system eventually declinesSetback
A fully implanted switch BCI remained functional over 7 years, but performance declined after roughly year 6.
UMC Utrecht BCI team
2025
Voice streams every eighty milliseconds
A 1,024-word ECoG system synthesized personalized speech in 80-ms increments at 47.5 words/minute in one stroke survivor.
UCSF and UC Berkeley
2025
Instant voice remains error-prone
Closed-loop neural voice synthesis was near-instant, but a neural-only evaluation still had 43.60% word error.
Maitreyee Wairagkar and colleagues
2026
Home use crosses 3,800 hoursDeployment
One man independently used a wired intracortical BCI for 3,801 hours over 19 months, producing 1.96 million words at 56 words/minute.
UC Davis and BrainGate
2026
Wireless speech trial implants firstWe are here
Connect-One implanted its first participant with a fully internal 421-electrode BCI designed to test text and synthesized speech.
Paradromics and University of Michigan
2027
Two-person safety endpoint scheduledExperimentTarget
Connect-One lists May 2027 primary completion for an estimated two participants; the primary outcome is device-related adverse events.
Paradromics Connect-One trial
2032
Six-year follow-up scheduledExperimentTarget
The registry estimates study completion in January 2032 after up to six years of long-term safety observation.
Paradromics Connect-One trial
2036-20361 event0 shown

Additional records

Scheduled target: “Ten-year wireless durability target”.

2036
Ten-year wireless durability targetDeploymentTarget
Paradromics designs Connexus for more than 10 years of stable recording; this company goal is unproved, not a scheduled approval.
Paradromics
— end of record · 24 events shown —
04 · The dataThe data

The data behind the verdictThe data behind the verdict

Read each measure with its units, conditions, uncertainty and source; targets remain distinct from observations.Read each chart with its conditions, limits and source. Future targets remain distinct from recorded results.

Open figure in archive: Raw speed rose sharply, but the points mix laboratory prompts, voice synthesis and real-world personal use; they must be paired with error and vocabulary.Open figure in archive: Text-decoding error has fallen in some participants; the 2025 value evaluates direct voice intelligibility and is not comparable to text WER.Open figure in archive: Large-vocabulary text is demonstrated, but voice systems still often use 1,024-word constrained evaluations and language models shape output.Open figure in archive: Use moved from hundreds of supervised conversation hours to thousands of independently initiated home hours in one person.Open figure in archive: The current strongest deployment evidence is one participant using the system on 444 of 653 observed days, not a multi-user clinical cohort.Open figure in archive: The 2023 large-vocabulary speech system more than tripled the prior 18-word-per-minute handwriting BCI rate.Open figure in archive: Benchmark accuracy was much higher than the strict real-world rate of sentences immediately decoded perfectly.Open figure in archive: Independent use expanded more than fifteenfold between the 2024 and 2026 reports.Open figure in archive: The 2026 home system required a substantial daily setup despite independent operation afterward.
Error-adjusted decoded communication rateError-adjusted decoded communication rate
0408010014325
78 WPM × 75% correct; 1,024-word prompted task58.5 correct words/minute · 2023 · 7 moprompted laboratory speech task; avatar output
  1. 115.2 WPM × 74.4% correct; 50-word lab task · 11.31 correct words/minute · 2021 · prompted laboratory task; handwriting or 50-word vocabulary
  2. 247.5 WPM × 56.4% correct; direct synthesized-voice evaluation · 26.79 correct words/minute · 2025 · direct synthesized-voice evaluation
  3. 332 WPM × 97.5% correct; self-paced conversations · 31.2 correct words/minute · 2024 · self-paced conversational study
  4. 462 WPM × 76.2% correct; 125,000-word prompted task · 47.24 correct words/minute · 2023 · 7 mo · prompted laboratory speech task; large vocabulary
  5. 549.7 WPM × 96.5% correct; silent-speech benchmark after day 600 · 47.96 correct words/minute · 2026 · silent-speech benchmark after prolonged home use
  1. ASustained independent open conversation at the low end of natural speech · 120 correct words/minuteSustained independent open conversation at the low end of natural speech · 120 correct words/minute

Records are separate measurements; no line joins them. Values multiply reported words per minute by one minus word-error rate; this derived proxy prevents speed records from hiding errors. Tasks are not directly comparable: vocabularies, implants, language models, prompted versus spontaneous speech and text versus voice differ. The candidate file's recommended endpoint is error-free independent home communication; only the 2026 participant demonstrated prolonged independent home use, and its 47.96 point is still a structured benchmark rather than audited free-conversation throughput.

View data table
All independent metric records
DateValueStatusCondition / labelVerificationCICensoringSource
202111.31 correct words/minuteobserved15.2 WPM × 74.4% correct; 50-word lab task
prompted laboratory task; handwriting or 50-word vocabulary		unspecified—none2021 anarthria speech neuroprosthesis
2023.647.24 correct words/minuteobserved62 WPM × 76.2% correct; 125,000-word prompted task
prompted laboratory speech task; large vocabulary		unspecified—none2023 large-vocabulary intracortical speech BCI
2023.6158.5 correct words/minuteobserved78 WPM × 75% correct; 1,024-word prompted task
prompted laboratory speech task; avatar output		unspecified—none2023 ECoG speech and avatar neuroprosthesis
202431.2 correct words/minuteobserved32 WPM × 97.5% correct; self-paced conversations
self-paced conversational study		unspecified—none2024 rapidly calibrating speech neuroprosthesis
202526.79 correct words/minuteobserved47.5 WPM × 56.4% correct; direct synthesized-voice evaluation
direct synthesized-voice evaluation		unspecified—none2025 instantaneous voice synthesis
202647.96 correct words/minuteobserved49.7 WPM × 96.5% correct; silent-speech benchmark after day 600
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
Metric context · these are authored references, not observed records
ContextValueLabelBasisSource
Goal120 correct words/minuteSustained independent open conversation at the low end of natural speech—2021 anarthria speech neuroprosthesis · 2023 large-vocabulary intracortical speech BCI · 2023 ECoG speech and avatar neuroprosthesis · 2024 rapidly calibrating speech neuroprosthesis · 2025 instantaneous voice synthesis · 2026 long-term independent home use
AS OF July 2026
View every data figure
Error-adjusted decoded communication rateError-adjusted decoded communication rate
0408010014325
78 WPM × 75% correct; 1,024-word prompted task58.5 correct words/minute · 2023 · 7 moprompted laboratory speech task; avatar output
  1. 115.2 WPM × 74.4% correct; 50-word lab task · 11.31 correct words/minute · 2021 · prompted laboratory task; handwriting or 50-word vocabulary
  2. 247.5 WPM × 56.4% correct; direct synthesized-voice evaluation · 26.79 correct words/minute · 2025 · direct synthesized-voice evaluation
  3. 332 WPM × 97.5% correct; self-paced conversations · 31.2 correct words/minute · 2024 · self-paced conversational study
  4. 462 WPM × 76.2% correct; 125,000-word prompted task · 47.24 correct words/minute · 2023 · 7 mo · prompted laboratory speech task; large vocabulary
  5. 549.7 WPM × 96.5% correct; silent-speech benchmark after day 600 · 47.96 correct words/minute · 2026 · silent-speech benchmark after prolonged home use
  1. ASustained independent open conversation at the low end of natural speech · 120 correct words/minuteSustained independent open conversation at the low end of natural speech · 120 correct words/minute

Records are separate measurements; no line joins them. Values multiply reported words per minute by one minus word-error rate; this derived proxy prevents speed records from hiding errors. Tasks are not directly comparable: vocabularies, implants, language models, prompted versus spontaneous speech and text versus voice differ. The candidate file's recommended endpoint is error-free independent home communication; only the 2026 participant demonstrated prolonged independent home use, and its 47.96 point is still a structured benchmark rather than audited free-conversation throughput.

View data table
All independent metric records
DateValueStatusCondition / labelVerificationCICensoringSource
202111.31 correct words/minuteobserved15.2 WPM × 74.4% correct; 50-word lab task
prompted laboratory task; handwriting or 50-word vocabulary		unspecified—none2021 anarthria speech neuroprosthesis
2023.647.24 correct words/minuteobserved62 WPM × 76.2% correct; 125,000-word prompted task
prompted laboratory speech task; large vocabulary		unspecified—none2023 large-vocabulary intracortical speech BCI
2023.6158.5 correct words/minuteobserved78 WPM × 75% correct; 1,024-word prompted task
prompted laboratory speech task; avatar output		unspecified—none2023 ECoG speech and avatar neuroprosthesis
202431.2 correct words/minuteobserved32 WPM × 97.5% correct; self-paced conversations
self-paced conversational study		unspecified—none2024 rapidly calibrating speech neuroprosthesis
202526.79 correct words/minuteobserved47.5 WPM × 56.4% correct; direct synthesized-voice evaluation
direct synthesized-voice evaluation		unspecified—none2025 instantaneous voice synthesis
202647.96 correct words/minuteobserved49.7 WPM × 96.5% correct; silent-speech benchmark after day 600
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
Metric context · these are authored references, not observed records
ContextValueLabelBasisSource
Goal120 correct words/minuteSustained independent open conversation at the low end of natural speech—2021 anarthria speech neuroprosthesis · 2023 large-vocabulary intracortical speech BCI · 2023 ECoG speech and avatar neuroprosthesis · 2024 rapidly calibrating speech neuroprosthesis · 2025 instantaneous voice synthesis · 2026 long-term independent home use
AS OF July 2026
Raw speed rose sharply, but the points mix laboratory prompts, voice synthesis and real-world personal use; they must be paired with error and vocabulary.Raw speed rose sharply, but the points mix laboratory prompts, voice synthesis and real-world personal use; they must be paired with error and vocabulary.

Independent records

Reported decoding rate linear scale · words/minute

linear scale · words/minute
202115.2 words/minute

Conditionprompted laboratory task; handwriting or 50-word vocabulary

2023.662 words/minute

Conditionprompted laboratory speech task; large vocabulary

2023.6178 words/minute

Conditionprompted laboratory speech task; avatar output

202432 words/minute

Conditionself-paced conversational study

202547.5 words/minute

Conditiondirect synthesized-voice evaluation

202656 words/minute

Conditionsilent-speech benchmark after prolonged home use

observed self-reported target / pathway right-censored lower bound
View data table
All plotted observations
SeriesDateValueStatusCondition / labelVerificationCICensoringSource
Reported decoding rate202115.2 words/minuteobservedReported decoding rate
prompted laboratory task; handwriting or 50-word vocabulary		unspecified—none2021 anarthria speech neuroprosthesis
Reported decoding rate2023.662 words/minuteobservedReported decoding rate
prompted laboratory speech task; large vocabulary		unspecified—none2023 large-vocabulary intracortical speech BCI
Reported decoding rate2023.6178 words/minuteobservedReported decoding rate
prompted laboratory speech task; avatar output		unspecified—none2023 ECoG speech and avatar neuroprosthesis
Reported decoding rate202432 words/minuteobservedReported decoding rate
self-paced conversational study		unspecified—none2024 rapidly calibrating speech neuroprosthesis
Reported decoding rate202547.5 words/minuteobservedReported decoding rate
direct synthesized-voice evaluation		unspecified—none2025 streaming brain-to-voice neuroprosthesis
Reported decoding rate202656 words/minuteobservedReported decoding rate
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
AS OF July 2026
Text-decoding error has fallen in some participants; the 2025 value evaluates direct voice intelligibility and is not comparable to text WER.Text-decoding error has fallen in some participants; the 2025 value evaluates direct voice intelligibility and is not comparable to text WER.

Independent records

Reported word error rate linear scale · %

linear scale · %
202125.6 %

Conditionprompted laboratory task; handwriting or 50-word vocabulary

2023.623.8 %

Conditionprompted laboratory speech task; large vocabulary

2023.6125 %

Conditionprompted laboratory speech task; avatar output

20242.5 %

Conditionself-paced conversational study

202543.6 %

Conditiondirect synthesized-voice evaluation

20263.5 %

Conditionsilent-speech benchmark after prolonged home use

observed self-reported target / pathway right-censored lower bound
View data table
All plotted observations
SeriesDateValueStatusCondition / labelVerificationCICensoringSource
Reported word error rate202125.6 %observedReported word error rate
prompted laboratory task; handwriting or 50-word vocabulary		unspecified—none2021 anarthria speech neuroprosthesis
Reported word error rate2023.623.8 %observedReported word error rate
prompted laboratory speech task; large vocabulary		unspecified—none2023 large-vocabulary intracortical speech BCI
Reported word error rate2023.6125 %observedReported word error rate
prompted laboratory speech task; avatar output		unspecified—none2023 ECoG speech and avatar neuroprosthesis
Reported word error rate20242.5 %observedReported word error rate
self-paced conversational study		unspecified—none2024 rapidly calibrating speech neuroprosthesis
Reported word error rate202543.6 %observedReported word error rate
direct synthesized-voice evaluation		unspecified—none2025 instantaneous voice synthesis
Reported word error rate20263.5 %observedReported word error rate
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
AS OF July 2026
Large-vocabulary text is demonstrated, but voice systems still often use 1,024-word constrained evaluations and language models shape output.Large-vocabulary text is demonstrated, but voice systems still often use 1,024-word constrained evaluations and language models shape output.

Independent records

Decoder vocabulary size log scale · words

log scale · words
202150 words

Conditionprompted laboratory task; handwriting or 50-word vocabulary

2023.6125,000 words

Conditionprompted laboratory speech task; large vocabulary

2023.611,024 words

Conditionprompted laboratory speech task; avatar output

2024125,000 words

Conditionself-paced conversational study

20251,024 words

Conditiondirect synthesized-voice evaluation

2026125,000 words

Conditionsilent-speech benchmark after prolonged home use

observed self-reported target / pathway right-censored lower bound
View data table
All plotted observations
SeriesDateValueStatusCondition / labelVerificationCICensoringSource
Decoder vocabulary size202150 wordsobservedDecoder vocabulary size
prompted laboratory task; handwriting or 50-word vocabulary		unspecified—none2021 anarthria speech neuroprosthesis
Decoder vocabulary size2023.6125000 wordsobservedDecoder vocabulary size
prompted laboratory speech task; large vocabulary		unspecified—none2023 large-vocabulary intracortical speech BCI
Decoder vocabulary size2023.611024 wordsobservedDecoder vocabulary size
prompted laboratory speech task; avatar output		unspecified—none2023 ECoG speech and avatar neuroprosthesis
Decoder vocabulary size2024125000 wordsobservedDecoder vocabulary size
self-paced conversational study		unspecified—none2024 rapidly calibrating speech neuroprosthesis
Decoder vocabulary size20251024 wordsobservedDecoder vocabulary size
direct synthesized-voice evaluation		unspecified—none2025 streaming brain-to-voice neuroprosthesis
Decoder vocabulary size2026125000 wordsobservedDecoder vocabulary size
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
AS OF July 2026
Use moved from hundreds of supervised conversation hours to thousands of independently initiated home hours in one person.Use moved from hundreds of supervised conversation hours to thousands of independently initiated home hours in one person.

Independent records

Cumulative conversational or home-use exposure log scale · hours

log scale · hours
2024248 hours

Conditionself-paced conversational study

20263,801 hours

Conditionsilent-speech benchmark after prolonged home use

observed self-reported target / pathway right-censored lower bound
View data table
All plotted observations
SeriesDateValueStatusCondition / labelVerificationCICensoringSource
Cumulative conversational or home-use exposure2024248 hoursobservedCumulative conversational or home-use exposure
self-paced conversational study		unspecified—none2024 rapidly calibrating speech neuroprosthesis
Cumulative conversational or home-use exposure20263801 hoursobservedCumulative conversational or home-use exposure
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
AS OF July 2026
The current strongest deployment evidence is one participant using the system on 444 of 653 observed days, not a multi-user clinical cohort.The current strongest deployment evidence is one participant using the system on 444 of 653 observed days, not a multi-user clinical cohort.

Independent records

Independent home-use days linear scale · days used

linear scale · days used
2026444 days used

Conditionsilent-speech benchmark after prolonged home use

observed self-reported target / pathway right-censored lower bound
View data table
All plotted observations
SeriesDateValueStatusCondition / labelVerificationCICensoringSource
Independent home-use days2026444 days usedobservedIndependent home-use days
silent-speech benchmark after prolonged home use		unspecified—none2026 long-term independent home use
AS OF July 2026
The 2023 large-vocabulary speech system more than tripled the prior 18-word-per-minute handwriting BCI rate.The 2023 large-vocabulary speech system more than tripled the prior 18-word-per-minute handwriting BCI rate.

The 2023 large-vocabulary speech system more than tripled the prior 18-word-per-minute handwriting BCI rate.

2021 handwriting BCI18 words/minute equivalent
2023 speech decoder62 words/minute
Each value is scaled against zero in its own unit; bar length is not a cross-unit comparison.
View data table
All aligned comparisons
ClaimABSource / provenance
The 2023 large-vocabulary speech system more than tripled the prior 18-word-per-minute handwriting BCI rate.18 words/minute equivalent · 2021 handwriting BCI62 words/minute · 2023 speech decoder2023 large-vocabulary intracortical speech BCI
AS OF July 2026
Benchmark accuracy was much higher than the strict real-world rate of sentences immediately decoded perfectly.Benchmark accuracy was much higher than the strict real-world rate of sentences immediately decoded perfectly.

Benchmark accuracy was much higher than the strict real-world rate of sentences immediately decoded perfectly.

Prompted vocalized benchmark99.2 % word accuracy
Personal-use sentences immediately perfect53.3 % sentences
Each value is scaled against zero in its own unit; bar length is not a cross-unit comparison.
View data table
All aligned comparisons
ClaimABSource / provenance
Benchmark accuracy was much higher than the strict real-world rate of sentences immediately decoded perfectly.99.2 % word accuracy · Prompted vocalized benchmark53.3 % sentences · Personal-use sentences immediately perfect2026 long-term independent home use
AS OF July 2026
Independent use expanded more than fifteenfold between the 2024 and 2026 reports.Independent use expanded more than fifteenfold between the 2024 and 2026 reports.

Independent use expanded more than fifteenfold between the 2024 and 2026 reports.

2024 conversation exposure248 hours
2026 independent home use3,801 hours
Each value is scaled against zero in its own unit; bar length is not a cross-unit comparison.
View data table
All aligned comparisons
ClaimABSource / provenance
Independent use expanded more than fifteenfold between the 2024 and 2026 reports.248 hours · 2024 conversation exposure3801 hours · 2026 independent home use2026 long-term independent home use
AS OF July 2026
The 2026 home system required a substantial daily setup despite independent operation afterward.The 2026 home system required a substantial daily setup despite independent operation afterward.

The 2026 home system required a substantial daily setup despite independent operation afterward.

Typical setup20 minutes
Maximum continuous use19 hours
Each value is scaled against zero in its own unit; bar length is not a cross-unit comparison.
View data table
All aligned comparisons
ClaimABSource / provenance
The 2026 home system required a substantial daily setup despite independent operation afterward.20 minutes · Typical setup19 hours · Maximum continuous use2026 long-term independent home use
AS OF July 2026

Methods & conditions

Under the measurement.

The boundary of a measurement changes what it means.

Technical note 1 of 8

Online bilingual decoding was windowed

Method and conditions

The bilingual system classified 104 bilingual words with median cross-validated neural-only accuracy of 47.24%; language models then assembled sentences. It operated in successive 3.5-second windows, so “online” did not mean continuously streaming speech.

Source notes
Nature Biomedical Engineering

Technical notes and sources

Each note keeps its authored method, conditions and citations available for review.

Online bilingual decoding was windowed

The bilingual system classified 104 bilingual words with median cross-validated neural-only accuracy of 47.24%; language models then assembled sentences. It operated in successive 3.5-second windows, so “online” did not mean continuously streaming speech.

Nature Biomedical Engineering
Bilingual transfer remained single-participant

Bilingual transfer learning let neural data from one language improve decoding of the other, consistent with shared articulatory representations. Performance remained stable across a 30-day recording break without retraining, but this was still one participant and a constrained vocabulary.

Nature Biomedical Engineering
Inner speech is not unrestricted thought reading

Inner speech produced weaker but correlated motor-cortex patterns relative to attempted speech. Real-time decoding was demonstrated in three participants, using both 50-word and 125,000-word regimes; this was instructed, self-paced imagery, not unrestricted thought reading.

Cell
A neural lock can suppress unintended output

The inner-speech study identified a neural motor-intent dimension separating imagined from attempted speech. An “imagery-silenced” training strategy suppressed unintended output, while a mental-keyword lock achieved 98.75% activation accuracy in one participant.

Cell
Loudness intent appears before speech

In the 2026 loudness study, four attempted modes-mimed, whispered, normal and loud-were decoded offline with 94% and 89% accuracy in two participants. Intent-related preparatory signals crossed 80% accuracy 640 ms and 270 ms before speech onset, respectively.

Nature Communications
Buffer cadence is not response latency

The 2025 streaming system’s 80-ms increments are audio-buffer cadence, not total response latency. Median cue-to-output onset was 1.67 seconds for synthesized speech, and direct speech transcription had 40.8% median word error rate on its 1,024-word general set.

Nature Neuroscience
Instantaneous output still had high error

The “instantaneous” 2025 intracortical system decoded intonation and short pitch sequences, but neural-only synthesized speech still had 43.60% word error rate. Simultaneous residual-speech biosignals did not explain the result: stethoscopic-microphone decoding had 100% word error rate in the reported comparison.

Nature
Cross-participant transfer is not assistive use

Cross-participant transfer learning improved phoneme decoding across 25 speaking epilepsy patients and more than 3,600 sEEG electrodes. This is useful evidence for scalable model initialization, but it was not an assistive speech system demonstrated in an anarthric user.

Nature Communications
05 · What it unlocksWhat it unlocks

If the remaining tests pass.If the remaining tests pass.

Downstream capabilities remain conditional on results that are not yet in.These are possible next capabilities, conditional on results that are not yet in.

TODAY
  1. PASSED · Aug 202301 ↗
    Say more than fixed phrasesOpen-vocabulary brain-to-textSystems now decode from about 125,000 words in individual participants.Decode unconstrained attempted speech from a vocabulary large enough for ordinary English at useful speed.
  2. OPEN TEST02 ↗
    Restore a natural voiceReal-time expressive voiceStreaming and instantaneous systems exist, but direct neural voice still had 43.6% WER in one 2025 evaluation.Synthesize intelligible personalized speech with conversational latency, prosody and unconstrained vocabulary.
  3. OPEN TEST03 ↗
    Work every day at homeIndependent multi-year home useOne participant logged 3,801 hours over 19 months, with care partners still handling connections and startup.Operate at home for years with little calibration or researcher support and stable useful communication.
  4. OPEN TEST04 ↗
    Work for many peopleReproducible clinical benefitCurrent speech records are dominated by single-participant reports and investigational implants.Demonstrate safe, durable benefit across dozens of users and disease anatomies in a regulated multicenter trial.
PASSED · Aug 202301 ↗
Say more than fixed phrasesOpen-vocabulary brain-to-textSystems now decode from about 125,000 words in individual participants.Decode unconstrained attempted speech from a vocabulary large enough for ordinary English at useful speed.
TODAY
OPEN TEST02 ↗
Restore a natural voiceReal-time expressive voiceStreaming and instantaneous systems exist, but direct neural voice still had 43.6% WER in one 2025 evaluation.Synthesize intelligible personalized speech with conversational latency, prosody and unconstrained vocabulary.
OPEN TEST03 ↗
Work every day at homeIndependent multi-year home useOne participant logged 3,801 hours over 19 months, with care partners still handling connections and startup.Operate at home for years with little calibration or researcher support and stable useful communication.
OPEN TEST04 ↗
Work for many peopleReproducible clinical benefitCurrent speech records are dominated by single-participant reports and investigational implants.Demonstrate safe, durable benefit across dozens of users and disease anatomies in a regulated multicenter trial.

Possible, if the open tests pass.

  1. Restored conversation after paralysis

    This is a possible benefit only if restored conversation after paralysis works safely, reliably and at useful scale.Potential benefit only if the stated use (Restored conversation after paralysis) is demonstrated with evidence for safety, reliability, affordability and scale.

  2. Independent digital work

    This is a possible benefit only if independent digital work works safely, reliably and at useful scale.Potential benefit only if the stated use (Independent digital work) is demonstrated with evidence for safety, reliability, affordability and scale.

  3. Richer neural prostheses

    This is a possible benefit only if richer neural prostheses works safely, reliably and at useful scale.Potential benefit only if the stated use (Richer neural prostheses) is demonstrated with evidence for safety, reliability, affordability and scale.

These are possible outcomes, if the open tests pass. They are conditional, not promises.
06 · The players & the moneyThe players & the money

Meet the builders.Meet the builders.

The organisations, their approaches, and the evidence behind them.The teams doing the work, the money around them, and what the record can support.

USA · company

Paradromics

High-data-rate intracortical Connexus BCI for communication and computer control

Funding contextSeries A · $33M · May 2023Connect-One is an estimated two-person early-feasibility study, not an approved product.
Named investorsPrime Movers Lab · Westcott Investment Group · Dolby Family Ventures · Green Sands Equity

Capital is context, not a demonstration.

Company source ↗
View every builder record

Paradromics

USA · company · High-data-rate intracortical Connexus BCI for communication and computer control

Funding: Series A · $33M · May 2023

Named investors: Prime Movers Lab · Westcott Investment Group · Dolby Family Ventures · Green Sands Equity

Connect-One is an estimated two-person early-feasibility study, not an approved product.

Company source ↗

Precision Neuroscience

USA · company · Thin cortical-surface array intended to let people with paralysis control computers

Funding: Series C · $102M · Dec 2024

Named investors: General Equity Holdings · B Capital · Duquesne Family Office · Steadview Capital

The company reports total funding; a take-home speech product has not been demonstrated.

Company source ↗

Synchron

USA / Australia · company · Endovascular BCI for digital-device control by people with paralysis

Funding: Series C · $75M · Dec 2022

Named investors: ARCH Venture Partners · Gates Frontier · Bezos Expeditions · Khosla Ventures

Primarily a digital-control interface, not a high-bandwidth speech decoder.

Company source ↗

UCSF Chang Lab / BRAVO

USA · lab · Develops speech neuroprostheses and direct brain-to-voice systems for people with paralysis

Funding: Funding not recorded in this record.

Company source ↗

UC Davis Neuroprosthetics Lab

USA · lab · Clinical intracortical research to restore rapid text and personalized synthesized speech

Funding: Funding not recorded in this record.

Company source ↗

Utrecht CortiCom team

Netherlands · lab · 128-channel cortical communication study targeting independent home and keyword-speech use

Funding: Funding not recorded in this record.

Registered estimated enrollment: two participants.

Company source ↗
07 · SourcesSources

Follow the evidence.Follow the evidence.

40 sources — every figure on this page traces to one.40 sources — every figure on this page traces to one.

Source · what this page citesScope and limits
  1. 01

    Vidal's direct brain-computer communication proposal

    Read source ↗pubmed.ncbi.nlm.nih.gov
    Used on this page
    All 2 uses across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  2. 02

    Scope: Not recorded. Limit: Not recorded.

  3. 03

    2009 review of speech communication BCIs

    Read source ↗pmc.ncbi.nlm.nih.gov
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  4. 04

    2021 anarthria speech neuroprosthesis

    Read source ↗nejm.org
    Used on this page
    All 14 uses across 9 destinations

    Scope: Not recorded. Limit: Not recorded.

  5. 05

    2023 large-vocabulary intracortical speech BCI

    Read source ↗nature.com
    Used on this page
    All 13 uses across 8 destinations

    Scope: Primary publication named by 2023 large-vocabulary intracortical speech BCI. Limit: The publication's measured scope does not establish downstream engineering or deployment claims.

  6. 06

    2023 ECoG speech and avatar neuroprosthesis

    Read source ↗nature.com
    Used on this page
    All 10 uses across 6 destinations

    Scope: Primary publication named by 2023 ECoG speech and avatar neuroprosthesis. Limit: The publication's measured scope does not establish downstream engineering or deployment claims.

  7. 07

    2024 rapidly calibrating speech neuroprosthesis

    Read source ↗nejm.org
    Used on this page
    All 12 uses across 7 destinations

    Scope: Not recorded. Limit: Not recorded.

  8. 08

    Scope: Primary publication named by 2025 streaming brain-to-voice neuroprosthesis. Limit: The publication's measured scope does not establish downstream engineering or deployment claims.

  9. 09

    Scope: Primary publication named by 2025 instantaneous voice synthesis. Limit: The publication's measured scope does not establish downstream engineering or deployment claims.

  10. 10

    2026 long-term independent home use

    Read source ↗nature.com
    Used on this page
    All 20 uses across 11 destinations

    Scope: Primary publication named by 2026 long-term independent home use. Limit: The publication's measured scope does not establish downstream engineering or deployment claims.

  11. 11

    Scope: Not recorded. Limit: Not recorded.

  12. 12

    Paradromics Connect-One first implant

    Read source ↗uofmhealth.org
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  13. 13

    Scope: Not recorded. Limit: Not recorded.

  14. 14

    Timeline · BrainGate proves multidimensional control

    Read source ↗nature.com
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  15. 15

    Timeline · Implant drives vowel synthesizer

    Read source ↗isca-archive.org
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  16. 16

    Timeline · Cursor typing accelerates

    Read source ↗elifesciences.org
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  17. 17

    Timeline · Cortical activity synthesizes sentences

    Read source ↗pmc.ncbi.nlm.nih.gov
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  18. 18

    Timeline · Handwriting bypass reaches phone speed

    Read source ↗nature.com
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  19. 19

    Timeline · Bilingual speech is decoded

    Read source ↗nature.com
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  20. 20

    Timeline · Seven-year system eventually declines

    Read source ↗nejm.org
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  21. 21

    Timeline · Instructed inner speech decodes in real time

    Read source ↗pubmed.ncbi.nlm.nih.gov
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  22. 22

    Scope: Not recorded. Limit: Not recorded.

  23. 23

    Scope: Not recorded. Limit: Not recorded.

  24. 24

    Engineer note · Inner speech is not unrestricted thought reading

    Read source ↗pmc.ncbi.nlm.nih.gov

    Scope: Not recorded. Limit: Not recorded.

  25. 25

    Engineer note · Cross-participant transfer is not assistive use

    Read source ↗nature.com

    Scope: Not recorded. Limit: Not recorded.

  26. 26

    companies · Paradromics

    Read source ↗paradromics.com
    Used on this page
    All 2 uses across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  27. 27

    companies · Precision Neuroscience

    Read source ↗globenewswire.com
    Used on this page
    All 2 uses across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  28. 28

    companies · Synchron

    Read source ↗businesswire.com
    Used on this page
    All 2 uses across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  29. 29

    companies · UCSF Chang Lab / BRAVO

    Read source ↗changlab.ucsf.edu
    Used on this page
    All 1 use across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  30. 30

    companies · UC Davis Neuroprosthetics Lab

    Read source ↗health.ucdavis.edu
    Used on this page
    All 2 uses across 2 destinations

    Scope: Not recorded. Limit: Not recorded.

  31. 31

    Scope: Not recorded. Limit: Not recorded.

  32. 32

    people · Edward Chang

    Read source ↗neurosurgery.ucsf.edu
    Used on this page
    All 1 use across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  33. 33

    people · David Brandman

    Read source ↗giving.ucdavis.edu
    Used on this page
    All 1 use across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  34. 34

    people · Thomas Oxley

    Read source ↗research.unimelb.edu.au
    Used on this page
    All 1 use across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  35. 35

    Scope: Not recorded. Limit: Not recorded.

  36. 36

    geography · Australia

    Read source ↗medicine.unimelb.edu.au
    Used on this page
    All 1 use across 1 destination

    Scope: Not recorded. Limit: Not recorded.

  37. 37

    Paradromics official site

    Read source ↗paradromics.com

    Scope: Not recorded. Limit: Not recorded.

  38. 38

    Precision Neuroscience official site

    Read source ↗precisionneuro.com

    Scope: Not recorded. Limit: Not recorded.

  39. 39

    Synchron official site

    Read source ↗synchron.com

    Scope: Not recorded. Limit: Not recorded.

  40. 40

    UCSF Chang Lab / BRAVO official site

    Read source ↗changlab.ucsf.edu

    Scope: Not recorded. Limit: Not recorded.

Sources describe evidence. Inclusion is not endorsement.