name: data-and-features
description: "ASRT data configuration, datalist, dictionary, WAV, and speech
feature extraction operating guidance."
metadata:
disco-role: operating
disable-model-invocation: true
license: GPL 3.0
data-and-features
Use this sub-skill when an ASRT task involves dataset configuration, pinyin dictionaries, WAV metadata, or speech-feature extraction before model training, evaluation, prediction, or serving.
Owns
asrt_config.json structure and dataset section semantics.dict.txt tab-separated pinyin-to-symbol rows and pinyin-to-index mapping.- ASRT wav-list and syllable-label line schemas.
DataLoader data assembly behavior and utils.config cache behavior.- WAV readers and byte decoding used before feature extraction.
- MFCC, Logfbank, Spectrogram, and SpecAugment feature behavior, including the 16 kHz constraint for spectrogram-style inputs.
Route away
- Training, evaluation, prediction scripts, acoustic-model class selection, model weights, CTC, and tensor shapes beyond data/feature compatibility: use
acoustic-models.
- Pinyin sequence to Chinese text decoding or language-model internals: use
language-model.
- HTTP/gRPC request payloads, server/client execution, and SDK calls: use
serving-clients.
- Full dataset downloads and interactive download helpers: treat as reference-only, not as a bundled runtime workflow.
Operating path
- Read references/data-and-config.md before editing or validating config, dict, datalist, or syllable labels.
- Run scripts/validate_asrt_config.py for self-contained structural validation of config/list/dict files.
- Read references/audio-and-features.md before diagnosing WAV metadata, 16 kHz failures, spectrogram shape, MFCC/Logfbank behavior, or SpecAugment randomness.
- Run scripts/inspect_audio_features.py for safe WAV and feature-shape inspection without importing ASRT.
- Use references/api-reference.md for quick schema/API reminders and references/troubleshooting.md for failure diagnosis.
Safety and boundaries
The bundled scripts are read-only validators/inspectors. They do not download corpora, record microphone audio, call servers, train models, or require an ASRT checkout. Provide user-owned config, dict, datalist, label, or WAV paths explicitly.