Quiz dotyczący dostrajania szeptów

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Jakie zmiany zachodzą podczas nadzorowanego dostrajania Whisper?
   • Only the decoded transcript is edited after inference
   • Only the language label is changed without training
   • Model weights are updated using paired audio and text examples
   • Only a static list of spellings is appended to the output
2. Dlaczego segmenty audio i transkrypcje muszą być wyrównane?
   • Alignment guarantees the same GPU memory use for every segment
   • Misaligned text teaches the model incorrect speech-to-text mappings
   • An alignment error is harmless if the transcript contains domain vocabulary
   • Correct timing guarantees that every transcript word is accurate
3. Aby ocenić wydajność niewidocznych mówców, gdy każdy mówca wnosi wiele nagrań, jak należy podzielić dane?
   • Randomly split overlapping windows from each speaker
   • Place every speaker in every partition
   • Use test speakers during training only
   • Group by speaker before creating train, validation and test splits
4. Co wspiera wstrzymany zestaw walidacji podczas dostrajania?
   • Checkpoint and hyperparameter decisions
   • Unlimited tuning followed by an unbiased final score on that same set
   • Gradient updates using every held-out label
   • A guarantee that the final deployment distribution will be unchanged
5. Czym mogą różnić się raporty dotyczące poziomu błędów w dwóch słowach?
   • Only the number of model parameters
   • Only the loss observed on the training set
   • Text normalization and tokenization conventions
   • Only whether both systems use the same tokenizer name
6. Why monitor a general-domain set while adapting to a narrow domain?
   • General data always lowers target accuracy
   • Target adaptation can reduce performance outside its training distribution
   • The test set should be used for gradient updates
   • The model size changes automatically
7. What can low-rank adapters offer when the selected implementation supports them?
   • Guaranteed lower WER in every domain
   • Automatic correction of mislabeled transcripts
   • A guarantee that no general-domain behavior can change
   • Fewer trainable parameters during adaptation
8. Why compare fine-tuning with decoding adjustments or domain-term support?
   • Training may add cost when simpler changes solve the problem
   • Decoding choices always change model weights
   • Fine-tuning cannot affect vocabulary behavior
   • A baseline makes test labels unnecessary