Update TIMIT/stimuli/README.md

TIMIT/stimuli/README.md

@@ -0,0 +1,74 @@
+# TIMIT stimuli
+
+This directory documents the TIMIT speech stimuli used in LibriBrain100.
+
+The original TIMIT audio and associated source materials are **not
+redistributed in this repository** because they are third-party licensed
+material.
+
+## Source
+
+LibriBrain100 uses the TIMIT Acoustic-Phonetic Continuous Speech Corpus:
+
+- Linguistic Data Consortium catalogue: `LDC93S1`
+- 630 speakers
+- 8 major dialect regions of the United States
+- 10 sentences per speaker
+- 6,300 utterances in total
+- Original audio: 16-kHz, 16-bit linear PCM
+
+LibriBrain100 collected MEG recordings for the full set of 6,300 TIMIT
+utterances.
+
+Official corpus page:
+
+https://catalog.ldc.upenn.edu/LDC93S1
+
+## Obtaining the stimuli
+
+TIMIT is distributed by the Linguistic Data Consortium (LDC) under its own
+licence and is not redistributed as part of LibriBrain100.
+
+Users wishing to obtain the corresponding speech materials should acquire
+TIMIT directly from the LDC and comply with the applicable licence terms.
+
+## Stimulus preparation
+
+For presentation during the LibriBrain100 experiment, the source audio was
+converted to uncompressed WAV format and resampled from 16 kHz to 48 kHz
+using SoX.
+
+The experiment used all 6,300 TIMIT utterances. Linguistic annotations
+time-locked to the corresponding MEG recordings are provided in the
+LibriBrain100 event files.
+
+## Dataset split
+
+LibriBrain100 follows the standard TIMIT evaluation split where possible.
+
+For testing, we use the official 24-speaker TIMIT core test set, excluding
+the shared `SA1` and `SA2` sentences.
+
+For validation, we use the 50-speaker development set from the standard
+Kaldi TIMIT recipe, again excluding the shared `SA` sentences.
+
+Speakers are also excluded from training where necessary to prevent overlap
+with the test material. The complete validation and test speaker IDs are
+reported in the LibriBrain100 dataset paper.
+
+## Licence
+
+TIMIT is third-party licensed material and is not redistributed or
+relicensed as part of LibriBrain100.
+
+Users must obtain TIMIT independently from the Linguistic Data Consortium
+and comply with the applicable LDC licence.
+
+## Citation
+
+If you use these data, please cite the LibriBrain100 dataset paper and the
+original TIMIT corpus:
+
+Garofolo, J. S., Lamel, L. F., Fisher, W. M., Fiscus, J. G., Pallett,
+D. S., Dahlgren, N. L., and Zue, V. (1993). TIMIT Acoustic-Phonetic
+Continuous Speech Corpus. Linguistic Data Consortium, LDC93S1.