On the Role of Encoder Depth: Pruning Whisper and LoRA Fine-Tuning in SLAM-ASR

Dimensions overall score 7.0

• Pruning two encoder layers from the Whisper encoder in SLAM-ASR causes only 2-4% WER degradation.result 90%
  Evidencepartial

  pruning two encoder layers causes only 2-4% WER degradation

  Implicationpartial

  Directly stated in the abstract with specific numeric results across multiple languages and model variants.

  Verificationpartial

  partial

• Combining encoder layer pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7-14%.result 90%
  Evidencepartial

  combining this pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7-14%

  Implicationpartial

  Explicitly stated in the abstract with clear performance and parameter reduction metrics.

  Verificationpartial

  partial

• LoRA primarily compensates for pruning degradation through the language model's linguistic priors, reducing total word errors by 11-21% for Dutch and English.result 85%
  Evidencepartial

  LoRA primarily compensates through the language model's linguistic priors, reducing total word errors by 11-21% for Dutch and English

  Implicationpartial

  Directly stated in the abstract with specific error reduction percentages, though the exact mechanism is inferred from the analysis.

  Verificationpartial

  partial

• For low-resource Danish, LoRA's compensation is smaller (4-7%) and introduces increased insertion errors.limitation 90%
  Evidencepartial

  for low-resource Danish, the reduction is smaller (4-7%), and LoRA introduces increased insertion errors

  Implicationpartial

  Explicitly stated in the abstract with specific numeric results and a clear limitation.

  Verificationpartial

  partial

• The effectiveness of LoRA compensation for pruning depends on the LLM's pre-existing language proficiency and available training data.result 85%
  Evidencepartial

  compensation effectiveness depends on the LLM's pre-existing language proficiency and available training data

  Implicationpartial

  Strongly implied by the comparative results across languages with different resource levels, and stated as a conclusion in the abstract.

  Verificationpartial

  partial

• No prior work has examined whether LoRA can compensate for performance lost through structured encoder layer pruning in SLAM-ASR.technical 80%
  Evidencepartial

  no prior work has examined whether LoRA can compensate for performance lost through structured encoder layer pruning

  Implicationpartial

  Directly stated in the analysis section as a gap in prior research, though it is a claim about the novelty of the work.

  Verificationpartial

  partial

• Substitution and deletion errors show the largest reductions when LoRA is applied to compensate for pruning.result 80%
  Evidencepartial

  with substitutions and deletions showing the largest reductions

  Implicationpartial

  Directly stated in the abstract, though the evidence is part of a broader claim about error analysis.

  Verificationpartial

  partial

Competitive landscape

This research demonstrates a method to significantly reduce the parameter count of ASR models by pruning encoder layers and using LoRA fine-tuning, leading to performance improvements and efficiency gains.