Brain2voice 2.0: Intracortical Voice Synthesis BCI (Wairagkar et al., 2026)

Inputs

The measured or assumed values behind the calculations, each with its source.

rate = 25 word/min

Not reported by the paper as a WPM figure. Measured directly from the 22 benchmark-test-set target (attempted-speech-aligned) audio clips released with the paper: 149 words across 358.1 s = 24.96 word/min. Reflects participant T15's slow, largely one-word-at-a-time cadence from severe ALS dysarthria, distinct from his faster self-paced brain-to-text conversational rate (32–56 wpm; Card et al. 2024, 2026 — same participant, different task).

P = 0.9476

1 − 5.24% word error rate: median WER from naive human listeners transcribing the continuous-acoustic-head synthesized voice on benchmark test trials (this is the paper's primary evaluation output; Results §4.1, Table 1). The tokenized-acoustic-head output scored nearly identically (WER 5.65%).

H = 5.0 bits/word

Shannon per-word entropy of English, the same convention used across the atlas's other free-text/speech entries.

Strictest ITR

Each scoring method is an upper bound on the channel, so the headline is the strictest (smallest) one for this entry. Use the score selector on the home page to view any single method across entries.

What counts as a bit depends on the action space. The number of distinguishable actions and how likely each one is are design choices of the task, not the sensing hardware. The same modality can present a fixed set of targets, a set pruned per step by a grammar or language model, or a continuous control space. Each of these changes how many actions are live and how the probability mass is spread, and therefore the information per selection. Read the action space below before comparing headline numbers across entries.

Action space

What the user can produce at each step, and how those options are distributed.

Structure

Continuous control space

Size

Continuous

Prior

Context-conditioned: likelihoods depend on prior actions

Notes

Brain2voice 2.0 causally predicts continuous LPCNet acoustic features directly from intracortical activity every 10 ms and vocodes them to a waveform; there is no discrete word- or key-level classification step gating the primary output. (The model also emits an auxiliary phoneme stream — 39 phonemes + silence + blank — and RVQ acoustic tokens, but neither is the evaluated channel: intelligibility is scored by human listeners transcribing the synthesized voice.) With no natural fixed-N action space, a Wolpaw or achieved-bitrate selection bound doesn't soundly apply the way it does for keyboard- or vocabulary-list decoders; the Shannon word-entropy throughput is the only method used here. Evaluated on the same intracortical dataset (BrainGate2 participant T15, four 64-electrode arrays in ventral precentral gyrus) as the entry's predecessor, Wairagkar et al. 2025's instantaneous voice-synthesis neuroprosthesis (WER 43.75%); brain2voice 2.0 is an 8x intelligibility improvement over that prior SOTA on the same benchmark.

Comparability The strictest bound here is Fitts throughput: the index of difficulty, log₂(A/W + 1), per movement. Directly comparable to the other continuous-pointing entries (mouse, trackball, stylus, gaze and the cursor BCIs). Set against the text entries (keyboards, spellers, speech) it crosses methods: both report bits/s, but one measures movement difficulty and the other text information, so compare within the family first.

Source

Authors

Wairagkar, Srinivasan, Card, Brandman, Stavisky et al.

Publication

bioRxiv, 2026

Paper

10.64898/2026.06.30.735633

Reference

Wairagkar et al. 2026 (bioRxiv): 'Brain2voice 2.0: High-performance voice synthesis brain-computer interface'

Reference

Project page with synthesized voice examples

Reference

Wairagkar et al. 2025 (Nature): predecessor 'instantaneous voice-synthesis neuroprosthesis', WER 43.75%