Scaling Law in Neural Data: Non-Invasive Speech Decoding with 175 Hours of EEG Data
Paper • 2407.07595 • Published
Datasets:
textja list | texten list | transcript_ja stringlengths 1 239 | transcript_en stringlengths 3 695 | onset float32 0 10.9k | offset float32 -0.14 10.9k | session stringclasses 79
values | run stringclasses 8
values |
|---|---|---|---|---|---|---|---|
[
-0.000244140625,
-0.011474609375,
0.0211181640625,
0.036865234375,
-0.000743865966796875,
0.09423828125,
0.0205078125,
-0.0147705078125,
0.0255126953125,
0.0284423828125,
-0.0028533935546875,
-0.04638671875,
0.00518798828125,
-0.0255126953125,
-0.005340576171875,
-0.049072265625,
0.0... | [
-0.00046539306640625,
-0.04052734375,
0.0260009765625,
0.00665283203125,
-0.0021209716796875,
0.111328125,
0.09521484375,
-0.0023345947265625,
0.011474609375,
-0.01531982421875,
-0.007110595703125,
-0.0274658203125,
0.004241943359375,
-0.018310546875,
-0.01031494140625,
-0.0390625,
0... | ルーシーは伝統のスイッチをつけた | Lucy turned on the traditional switch. | 5 | 20.1 | ses-20230829 | run-01 |
[
-0.000244140625,
-0.011474609375,
0.0211181640625,
0.036865234375,
-0.000743865966796875,
0.09423828125,
0.0205078125,
-0.0147705078125,
0.0255126953125,
0.0284423828125,
-0.0028533935546875,
-0.04638671875,
0.00518798828125,
-0.0255126953125,
-0.005340576171875,
-0.049072265625,
0.0... | [
-0.00046539306640625,
-0.04052734375,
0.0260009765625,
0.00665283203125,
-0.0021209716796875,
0.111328125,
0.09521484375,
-0.0023345947265625,
0.011474609375,
-0.01531982421875,
-0.007110595703125,
-0.0274658203125,
0.004241943359375,
-0.018310546875,
-0.01031494140625,
-0.0390625,
0... | ルーシーは伝統のスイッチをつけた | Lucy turned on the traditional switch. | 10 | 25.1 | ses-20230829 | run-01 |
[
-0.000461578369140625,
0.0238037109375,
0.051025390625,
0.058837890625,
-0.001983642578125,
0.007568359375,
0.0966796875,
-0.004791259765625,
0.0306396484375,
0.07666015625,
0.0498046875,
0.0038604736328125,
-0.00762939453125,
-0.08056640625,
0.055419921875,
-0.02685546875,
0.0415039... | [
-0.0003185272216796875,
0.061767578125,
0.019287109375,
0.051513671875,
-0.0016326904296875,
-0.03564453125,
0.10595703125,
-0.004150390625,
0.0191650390625,
0.047607421875,
0.036865234375,
0.036865234375,
-0.005767822265625,
-0.0247802734375,
0.049560546875,
-0.005706787109375,
0.01... | 作曲家の中では誰が一番好きですか | Who is your favorite composer? | 15 | 30.1 | ses-20230829 | run-01 |
[
-0.0002346038818359375,
-0.03173828125,
0.044189453125,
-0.042724609375,
-0.001007080078125,
0.0341796875,
0.0703125,
-0.032958984375,
0.028564453125,
0.006011962890625,
-0.04296875,
0.0269775390625,
0.00445556640625,
-0.025634765625,
0.037109375,
-0.0576171875,
-0.03271484375,
0.0... | [
0.00006151199340820312,
-0.06689453125,
0.0072021484375,
-0.04345703125,
0.0001316070556640625,
0.042236328125,
0.1142578125,
-0.045166015625,
0.0400390625,
-0.011962890625,
-0.03564453125,
0.0303955078125,
0.005157470703125,
-0.0179443359375,
0.06884765625,
-0.033203125,
-0.03417968... | 壁に書かれた文字を反読しようとした | I tried to decipher the characters written on the wall. | 40 | 55.099998 | ses-20230829 | run-01 |
[
-0.0002651214599609375,
-0.051025390625,
0.0693359375,
0.0257568359375,
-0.000885009765625,
0.10791015625,
0.025390625,
0.0203857421875,
0.01904296875,
0.04833984375,
0.0908203125,
0.0169677734375,
0.003021240234375,
-0.09130859375,
0.01104736328125,
0.040771484375,
0.03564453125,
... | [
-0.000156402587890625,
-0.02880859375,
0.05615234375,
-0.031494140625,
-0.000133514404296875,
0.08642578125,
0.0732421875,
0.041015625,
0.0263671875,
0.007720947265625,
0.078125,
0.0228271484375,
0.00830078125,
-0.0849609375,
0.01251220703125,
0.0216064453125,
0.02783203125,
0.0468... | 嵐のため彼女は定時に到着できなかった | The storm prevented her from arriving on time. | 60 | 75.099998 | ses-20230829 | run-01 |
[-0.0002651214599609375,-0.051025390625,0.0693359375,0.0257568359375,-0.000885009765625,0.1079101562(...TRUNCATED) | [-0.000156402587890625,-0.02880859375,0.05615234375,-0.031494140625,-0.000133514404296875,0.08642578(...TRUNCATED) | 嵐のため彼女は定時に到着できなかった | The storm prevented her from arriving on time. | 65 | 80.099998 | ses-20230829 | run-01 |
[-0.0002841949462890625,-0.00164031982421875,-0.02001953125,-0.02294921875,-0.00128936767578125,0.01(...TRUNCATED) | [-0.000354766845703125,0.06591796875,0.05078125,-0.06787109375,-0.0015106201171875,0.0693359375,0.06(...TRUNCATED) | 病気のため寺園は学校に行けませんでした | Terazono couldn't go to school because of illness. | 70 | 85.099998 | ses-20230829 | run-01 |
[-0.000453948974609375,-0.09765625,0.01080322265625,-0.0284423828125,-0.001556396484375,0.1005859375(...TRUNCATED) | [-0.0003185272216796875,-0.09765625,0.037353515625,-0.07666015625,-0.0009002685546875,0.0986328125,0(...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) | "The ballerinas around the wall are stretching their legs and toes to keep their muscles from stiffe(...TRUNCATED) | 85 | 100.099998 | ses-20230829 | run-01 |
[-0.000453948974609375,-0.09765625,0.01080322265625,-0.0284423828125,-0.001556396484375,0.1005859375(...TRUNCATED) | [-0.0003185272216796875,-0.09765625,0.037353515625,-0.07666015625,-0.0009002685546875,0.0986328125,0(...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) | "The ballerinas around the wall are stretching their legs and toes to keep their muscles from stiffe(...TRUNCATED) | 90 | 105.099998 | ses-20230829 | run-01 |
[-0.000499725341796875,0.01446533203125,0.07177734375,-0.019775390625,-0.00286865234375,0.0485839843(...TRUNCATED) | [-0.000675201416015625,0.0311279296875,0.06005859375,-0.0281982421875,-0.003021240234375,0.076171875(...TRUNCATED) | 寛大すぎることが彼の最大の欠点です | Being too generous is his greatest flaw. | 100 | 115.099998 | ses-20230829 | run-01 |
End of preview. Expand in Data Studio
ds007808 sub-01 speechopen/pangolin — Qwen3-Embedding text targets
Companion to tankalapavankalyan/ds007808-sub01-speechopen-pangolin-preprocessed.
For every one of the 153,796 5-second EEG windows in that dataset, this repo provides a
frozen text embedding of the window's transcript — the CLIP target for EEG→text decoding,
the text analogue of the w2v (wav2vec2-XLSR-53) speech target in the main dataset.
Built to extend the EEG→speech replication of Sato et al. 2024, "Scaling Law in Neural Data" (arXiv:2407.07595) to EEG→text, changing only the frozen target encoder (everything else — EEG windows, split, loss, metric — is held identical).
Schema (one row per window)
| column | type | description |
|---|---|---|
textja |
float16[1024] |
Qwen3-Embedding-4B sentence embedding of the Japanese transcript, MRL-truncated to 1024-d, L2-normalized |
texten |
float16[1024] |
same encoder, the English translation |
transcript_ja |
string | Japanese transcript (the spoken language) |
transcript_en |
string | English translation (value_en) |
onset |
float32 | window onset (s), EEG time |
offset |
float32 | aligned-audio onset (s) |
session, run |
string | recording identifiers |
Splits: train / valid / test in the same global chronological 80/10/10 order as the
main dataset — so row i of split S here corresponds to row i of split S there. You can also
join on (session, run, onset).
How the targets were produced
- Encoder:
Qwen/Qwen3-Embedding-4B(frozen), #1 multilingual MTEB. Native 2560-d output MRL-truncated to 1024-d (to match the wav2vec2 target dim) then L2-normalized. One model for both languages → JA and EN targets are directly comparable. - Japanese text = the per-window transcript (overlap-join of
overtevents over the window). - English text = the identical overlap-join over the events'
value_encolumn, validated to reproduce the Japanese join byte-for-byte (0 mismatches across all 372 runs). - Embeddings computed once per unique string (126,976 JA / 126,816 EN) and scattered back.
Cross-lingual sanity: mean cosine between a window's JA and EN target = 0.749 (random pair 0.323).
Intended use
Drop-in CLIP target for EEG→text retrieval: train an EEG encoder to predict textja (native
decoding) or texten (cross-lingual / translation), evaluate top-1/top-10 retrieval over 512
candidates. Note the subject spoke Japanese, so textja is the direct analogue of the speech
target; texten tests whether semantic content is decodable through translation.
License: CC0-1.0 (inherits ds007808 / OpenNeuro).
- Downloads last month
- 84