Tujia speech features from educational videos: pilot dataset (110 utterances)

This package shares the derived acoustic features and quality metrics for 110 Tujia utterances selected from the public educational video series "Gen Wo Xue Shuo Tujia Yu" (跟我学说土家语). It does not contain raw video, raw audio, screenshots, speaker images, or personally identifiable information. Raw media cannot be redistributed because of platform copyright restrictions. Processing summary: Audio was standardized to 16,000 Hz, mono, 16-bit PCM, peak amplitude 0.99. A first-order pre-emphasis filter H(z) = 1 - 0.97 z^-1 was applied. Spectral subtraction was used for noise reduction. Voice activity detection used short-time energy and zero-crossing rate (25 ms frames, 10 ms shift, Hamming window; noise floor from the lowest-energy frames; high/low energy thresholds of noise floor +8/+3 dB; ZCR threshold 0.25; gaps shorter than 150 ms merged; 200 ms margins). Quality metrics: speech frame ratio >= 30%, SNR >= 10 dB (capped at 60 dB), clipping ratio < 1%, spectral dynamic range reported descriptively, duration consistency 0.3-20 s. Feature extraction: F0 range 50-500 Hz (normalized cross-correlation); LPC order 18 with formant constraints of 150-5000 Hz and bandwidth < 400 Hz; MFCC with 13 coefficients, 26 Mel filters, and 2048-point FFT. Contents: acoustic feature data (F0, F1/F2, mean MFCC C1-C5), quality assessment metrics, processing parameters, data dictionary, and the MATLAB processing pipeline (scripts/). Related manuscript: "Acoustic analysis of Tujia speech from educational videos: A low-resource dataset study". Ethics and copyright: The source videos are public educational content posted for language teaching purposes. No raw audio/video, video URLs, or creator-identifying metadata are included. The data were collected from open-access platforms without direct interaction with human subjects.