A 1000-hour EEG-EMG-audio dataset of Japanese speech production
Paper • 2606.01264 • Published
Datasets:
Tasks:
Automatic Speech Recognition
Formats:
parquet
Languages:
Japanese
Size:
100K - 1M
ArXiv:
License:
eeg array 2D | w2v array 2D | audio audioduration (s) 5 5 | transcript stringlengths 1 131 | onset float32 0 3.42k | offset float32 10.7 3.44k | vad float32 0.2 1 | session stringclasses 3
values | run stringclasses 6
values |
|---|---|---|---|---|---|---|---|---|
[[0.2646484375,0.29296875,0.46142578125,0.25634765625,-0.342041015625,-0.47216796875,-0.161743164062(...TRUNCATED) | [[-2.83984375,2.501953125,-0.77734375,-4.1171875,7.7109375,-0.07196044921875,2.314453125,3.05078125,(...TRUNCATED) | ルーシーは伝統のスイッチをつけた | 5 | 20.1 | 0.3276 | ses-20230829 | run-01 | |
[[-2.02734375,-2.193359375,-1.7509765625,-0.95068359375,-0.37158203125,-0.143798828125,-0.2459716796(...TRUNCATED) | [[-2.931640625,-5.13671875,7.54296875,2.501953125,-1.482421875,-0.429443359375,1.236328125,3.0625,0.(...TRUNCATED) | ルーシーは伝統のスイッチをつけた | 10 | 25.1 | 0.2148 | ses-20230829 | run-01 | |
[[0.170166015625,0.1513671875,0.35546875,0.73828125,1.0244140625,0.99169921875,0.638671875,0.5004882(...TRUNCATED) | [[-2.46484375,2.0546875,-1.875,-5.921875,3.404296875,-0.69140625,2.986328125,1.5712890625,-1.3623046(...TRUNCATED) | 作曲家の中では誰が一番好きですか | 15 | 30.1 | 0.6712 | ses-20230829 | run-01 | |
[[-0.958984375,-1.12890625,-0.728515625,-0.55126953125,-0.818359375,-0.884765625,-0.89892578125,-0.6(...TRUNCATED) | [[-2.0546875,2.767578125,1.7373046875,-4.88671875,3.51171875,-0.278564453125,3.56640625,5.13671875,-(...TRUNCATED) | 壁に書かれた文字を反読しようとした | 40 | 55.099998 | 0.5892 | ses-20230829 | run-01 | |
[[-0.08477783203125,-0.20556640625,0.27197265625,0.480224609375,0.435546875,0.1925048828125,0.199462(...TRUNCATED) | [[-2.421875,2.283203125,0.45849609375,-7.21484375,1.265625,-0.0667724609375,3.5703125,3.677734375,-1(...TRUNCATED) | 嵐のため彼女は定時に到着できなかった | 60 | 75.099998 | 0.4732 | ses-20230829 | run-01 | |
[[0.2333984375,0.1751708984375,0.145751953125,0.0087432861328125,-0.0182952880859375,0.0213317871093(...TRUNCATED) | [[-2.546875,-0.34716796875,4.265625,0.94873046875,-0.29248046875,-2.109375,-0.27294921875,0.32519531(...TRUNCATED) | 嵐のため彼女は定時に到着できなかった | 65 | 80.099998 | 0.21 | ses-20230829 | run-01 | |
[[1.232421875,0.2900390625,0.2467041015625,0.453857421875,0.685546875,0.5263671875,0.348388671875,0.(...TRUNCATED) | [[-1.2548828125,2.173828125,-2.44921875,-6.640625,-0.533203125,0.1982421875,3.576171875,1.337890625,(...TRUNCATED) | 病気のため寺園は学校に行けませんでした | 70 | 85.099998 | 0.6768 | ses-20230829 | run-01 | |
[[-0.9033203125,-1.1474609375,-1.3671875,-0.85107421875,-0.335693359375,-0.290283203125,-0.719726562(...TRUNCATED) | [[0.211181640625,2.814453125,0.6015625,-4.359375,6.7734375,-2.34375,2.83203125,3.052734375,-3.447265(...TRUNCATED) | "壁の周りのバレリーナたちは筋肉が硬直しないように足と足先を伸ばして(...TRUNCATED) | 85 | 100.099998 | 0.5772 | ses-20230829 | run-01 | |
[[-0.47705078125,-0.99658203125,-0.18115234375,0.28662109375,0.25146484375,0.2022705078125,0.4943847(...TRUNCATED) | [[-2.8984375,0.70556640625,5.15625,-0.58251953125,-2.25390625,-5.21875,-1.8056640625,3.79296875,-3.6(...TRUNCATED) | "壁の周りのバレリーナたちは筋肉が硬直しないように足と足先を伸ばして(...TRUNCATED) | 90 | 105.099998 | 0.8092 | ses-20230829 | run-01 | |
[[1.4033203125,1.1923828125,1.150390625,0.99365234375,0.55126953125,-0.24267578125,-1.193359375,-1.7(...TRUNCATED) | [[-4.875,2.69921875,2.904296875,0.2017822265625,-7.66796875,-3.248046875,0.12646484375,0.982421875,-(...TRUNCATED) | 寛大すぎることが彼の最大の欠点です | 100 | 115.099998 | 0.6588 | ses-20230829 | run-01 |
End of preview. Expand in Data Studio
ds007808 sub-01 / speechopen / pangolin — preprocessed EEG↔speech windows
Ready-to-train EEG↔speech windows for replicating the scaling experiment of Sato et al. 2024, "Scaling Law in Neural Data: Non-Invasive Speech Decoding with 175 Hours of EEG Data" (arXiv:2407.07595), built from the public ds007808 dataset (arXiv:2606.01264).
Slice = subject sub-01, task speechopen (overt speech), device pangolin (128-ch g.Pangolin) — the rig matching the 175 h paper. Each example is one 5-second window.
Contents (per window)
| field | shape / type | description |
|---|---|---|
eeg |
float16 [128, 6000] |
128-ch EEG @1200 Hz, NLMS-cleaned, z-scored along time, clipped ±5 |
w2v |
float16 [249, 1024] |
frozen wav2vec2-large-xlsr-53 latents (layers 14–18 averaged) |
audio |
16 kHz mono | the aligned 5 s of spoken audio |
transcript |
string | Japanese transcript overlapping the window |
onset,offset |
float | EEG-time onset (s) and audio-time offset (s) of the window |
vad |
float | fraction of speech in the window (Silero VAD) |
session,run |
string | provenance |
Splits: chronological 80/10/10 (train = earliest, test = latest), by global session/run/onset order — held-out test is future data, as in the paper.
Preprocessing (faithful to Sato et al. 2024)
acquire 1200 Hz → MNE denoise → NLMS adaptive filter (instantaneous; refs = EOG + upper/lower
orbicularis-oris EMG; μ=0.1, ε=1e-3) → 5-s non-overlapping windows → per-window z-score along
time + clip ±5 → keep windows with >20 % speech (Silero VAD). Audio 48→16 kHz; latents from
frozen wav2vec2. EEG↔audio alignment uses the per-run constant wav_onset − onset offset
(exact in this dataset).
Assumptions where the paper is silent (documented)
- MNE denoise cutoffs (Fig 2a not given): 50 Hz notch (+harmonics) + 0.5 Hz high-pass.
- wav2vec2 checkpoint/layers:
xlsr-53, layers 14–18 averaged. - EEG kept at the acquisition rate 1200 Hz (no downsample stated by the paper).
Load
pip install datasets librosa soundfile # librosa/soundfile decode the audio column
from datasets import load_dataset
ds = load_dataset("tankalapavankalyan/ds007808-sub01-speechopen-pangolin-preprocessed")
x = ds["train"][0]
eeg = x["eeg"] # [128, 6000] float16
w2v = x["w2v"] # [249, 1024] float16 (CLIP target)
audio = x["audio"]["array"]
For a CLIP retrieval model (paper): encode eeg (e.g. HTNet+Conformer) → cosine-align to w2v,
zero-shot top-1/top-10 over 512 candidates.
License & citation
CC0-1.0 (derived from ds007808, CC0). Please cite ds007808 (arXiv:2606.01264) and Sato et al. 2024 (arXiv:2407.07595). Preprocessing code: this dataset's companion repo (HTNet+Conformer CLIP pipeline).
- Downloads last month
- 686