Dynamic Superb Eval

dynamic-superb-eval

Dynamic-SUPERB: Towards A Dynamic, Collaborative, and Comprehensive Instruction-Tuning Benchmark for Speech — Huang et al. (2023) (arXiv:2309.09510, 2023)

What this evaluates

Evaluates instruction-tuned speech models on their ability to perform diverse speech and audio tasks using natural language instructions. It probes zero-shot generalization by testing performance on seen versus unseen tasks and instructions across six dimensions: content, speaker, semantics, degradation, paralinguistics, and audio.

Datasets

• Dynamic-SUPERB — total 55; splits: test (55); repo https://github.com/dynamic-superb/dynamic-superb

Metrics

• accuracy (primary) — range: percent
  • Standard classification accuracy: the percentage of instances where the model's generated output exactly matches the ground-truth label among the provided options.

Input / output format

Input: Audio waveform paired with a natural language instruction that specifies the task and provides a set of candidate answer options.

Output: A generated text string corresponding to one of the candidate options provided in the instruction.

Scoring recipe

correct = 0
for pred, gold in zip(predictions, gold_labels):
    if pred.strip().lower() == gold.strip().lower():
        correct += 1
accuracy = (correct / len(predictions)) * 100

Common pitfalls

• Models may generate outputs not present in the candidate options, causing accuracy to drop below the random baseline.
• Models often rely on superficial instruction patterns (bag-of-words) rather than semantic understanding, leading to poor generalization on unseen tasks.
• The random baseline is constructed by sampling from the label distribution of each task, not a uniform random guess.

Evidence (verbatim from paper)

Table 2 presents the evaluation results for each baseline on seen tasks in Dynamic-SUPERB. Due to space constraints, we report the average accuracy for each dimension.

Citation

@misc{huang2023dynamicsuperb,
  title={Dynamic-SUPERB: Towards A Dynamic, Collaborative, and Comprehensive Instruction-Tuning Benchmark for Speech},
  author={Huang et al. (2023)},
  year={2023},
  note={arXiv:2309.09510}
}

• arXiv: 2309.09510