The package

pip install "hit-asr[paper] @ git+https://github.com/huseyin-karaca/hit-asr" installs two packages.

hitasr — the paper

Module	What
configs	every configuration the paper reports, fixed and named: MAIN[<corpus>], the spaces of Table 2 (SPACES), ABLATION, TIMING, SYNTHETIC, REBUILD
experiment	the main experiment on one corpus (MainExperiment): the search, the 5×2 cross-validation, Tables 1, 2, 3 and 6, the record
paper	the manuscript's tables from the records, shown and written as LaTeX (ManuscriptTables)
ablation, timing, synthetic	the ablation, the cost measurement on raw audio, the synthetic regime-switch check
records	the records the notebooks read from the Hub
core	the corpus registry (DatasetSpec): splits, trio, licence; the Hub repositories and the local rebuild folder (rebuild_dir, use_hub)
hub	HitHub: the dataset repository named the active corpus's way; the audio bases
scoring	the WER scorer (Whisper's English normaliser)
extractor, models.*	one extractor per expert: decode, word-error counters, frame-level encoder states; the rebuild (python -m hitasr.extractor <corpus>) and its check against the published labels
labels, frames, store	the labels as an error matrix, the frame shards memory-mapped or on the GPU, the two aligned (RouterStore)
routers, training	the HIT-ASR router and the MLP-pool router; the objective and the training loop
arms	every model as fit / predict, its search space, the registry MODELS
rover	the transcript-fusion baselines: confidence-weighted ROVER, CN-MBR
tuning	the hold-out search (hpt) and the names its studies and folds are stored under
crossval	the routing metrics, one fold of every model, Table 3 and the routing behaviour
eda, ensemble	the dataset table, the oracle arithmetic of a group of experts
deploy, latency, bases	the router on raw audio, the cost measurements, the builder of the AfriSpeech-200 audio base

labkit — the experiment machinery

Independent of speech; HIT-ASR is built on it.

Module	What
hub	Hugging Face dataset repositories as a cache: one listing and one parallel fetch per step, every Hub error retried and never taken for a missing file, a read-only mode; a local folder (local:/path) in the same layout stands in for a repository
runlog	a content-addressed result cache: a step's result stored under a hash of what it depends on, per model and seed; a strict mode that never computes
studies	Optuna studies stored as trial ledgers on the Hub
search	search spaces as data (Param) and as source
cv	the 5×2 splits (and one repetition of them by seed), k-fold, the hold-out partition
significance	the paired test of Table 6 on the 5×2 folds (see the paper) and its table
pretty, env	tables in a notebook and in a terminal; seeds and the GPU