The notebooks
| Notebook |
What it reproduces |
Open |
main_ami_sdm |
Tables 1, 2, 3 and 6 on AMI (single distant microphone) |
 |
main_earnings22 |
the same on Earnings-22 |
|
main_peoples_speech |
the same on People's Speech |
|
main_afrispeech |
the same on AfriSpeech-200 |
|
ablation |
the design switches of HIT-ASR, and the number of experts |
|
timing |
the cost on raw audio |
|
synthetic |
the synthetic regime-switch check |
|
manuscript_tables |
every table of the paper, as LaTeX in tables/ |
|
extract |
the experts' labels and frames, rebuilt from the audio (level 3) |
|
Every configuration is fixed in hitasr.configs and imported by its notebook; the notebooks hold the calls and the
results, the package holds the code.
A main notebook, section by section
| § |
What it does |
Shows |
| — |
MainExperiment(MAIN["<corpus>"], level=LEVEL): the corpus, its expert trio, the protocol and the seeds |
|
| 1 |
the hyperparameter spaces of every model (hitasr.configs.SPACES) |
Table 2 |
| 2 |
the three experts' labels (and, from level 2, their frames); a check that the stored word-error counters reproduce from the stored transcripts |
|
| 3 |
the search: a random 25 % of the clips is held out, and every model's hyperparameters are chosen there by 10 random draws, each trained on 70 % of the hold-out and scored on the rest |
Table 1, the selected configurations |
| 4 |
the 5×2 cross-validation on the other 75 %: each of five repetitions halves the clips with its seed, every model is trained on one half and tested on the other, then the other way round |
|
| 5 |
every model's metrics as mean ± sd over the ten folds; the routers' selection distributions |
Table 3, routing behaviour |
| 6 |
HIT-ASR against every other system on the same folds, with a custom statistical test — for full details, please see the paper |
Table 6 |
| 7 |
the record: everything above in one JSON file |
|
The models
| Model |
What it is |
| HIT-ASR |
the hierarchical transformer router on the experts' frame-level encoder states |
| MLP-pool |
a router on the experts' mean-pooled encoder states |
| ADASTT |
gradient-boosted trees on the pooled states, trained with cross-entropy |
| ROVER (confidence-weighted), CN-MBR |
transcript-level fusion of the three experts' outputs |
| Best single expert, Random, Oracle |
the references: the best expert on the training half, a uniform pick, the per-clip best |
MLP-pool (hard CE), MLP-pool trained with hard cross-entropy only, is searched and trained as a control; it is kept in
each record and in the ablation notebook, and is not a row of Tables 3 and 6.