acceptodds
Under review as a conference paper at ICLR 2027

Cumulative Acoustic Variation for Training-Free ASR Data Pruning

Abstract

Training-data pruning for automatic speech recognition (ASR) aims to reduce training-data requirements, yet model-dependent selection criteria can incur sub- stantial computational overhead. To address this, we propose a model-free acous- tic pruning approach that ranks training utterances using cumulative statistics computed directly from speech signals. Specifically, we develop three utterance- level criteria based on accumulated spectral entropy, spectral flux, and log-Mel temporal variation, requiring neither pruning-specific model training nor learned- model inference. Under a fixed utterance-count budget, these criteria retain both the extent and the local acoustic characteristics of each utterance through cumula- tive scoring. Experiments with Conformer on LibriSpeech show that the proposed acoustic criteria achieve the best performance at 5% and 1% retention and re- main competitive with the best-performing Teacher–Student criterion at 30% and 10% retention, while reducing data-selection time from 13.73 hours to less than 0.5 hours without requiring GPUs. In particular, Log-Mel Temporal Variation achieves the lowest WER at both 5% and 1% retention; at 5%, it reduces WER relative to random selection by 26.1% and 15.3% on test-clean and test-other, re- spectively. These results demonstrate that simple cumulative acoustic statistics provide an efficient and effective alternative to model-dependent criteria for ASR training-data pruning, with especially strong performance under highly restrictive retention budgets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Automatic Speech Recognition, Data Pruning, Data Selection, Training Data Efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3HSQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free acoustic data-pruning method for ASR using cumulative spectral entropy, spectral flux, and log-Mel temporal variation to rank training utterances. Experiments with Conformer on LibriSpeech show that Log-Mel Temporal Variation achieves the best performance among evaluated methods at 5% and 1% retention, while reducing selection time from 13.73 hours for Teacher–Student scoring to approximately 0.42 hours. Ablations support cumulative over mean aggregation under an utterance-count budget. However, evaluation is limited to one dataset and model architecture, and the contribution of acoustic variation beyond utterance duration remains insufficiently isolated.

Critical strengths
  1. Excellent selection efficiency. The proposed criteria require neither pruning-specific model training nor learned-model inference. Log-Mel Temporal Variation scores the LibriSpeech training set in approximately 0.42 hours using CPUs, compared with 13.73 hours for Teacher–Student selection. This makes the approach attractive when computational resources are limited.
  2. Strong results under restrictive data budgets. At 5% retention, Log-Mel Temporal Variation achieves 10.5% WER on test-clean and 21.0% on test-other, outperforming the evaluated Teacher criterion (13.6% and 24.3%, respectively). The paper also compares cumulative and mean aggregation and includes simple word-count and phone-count baselines, providing useful evidence for its design choices.
Critical weaknesses
  1. Limited evaluation scope and statistical evidence. All experiments use LibriSpeech and a single Conformer training setup. The paper does not establish whether the gains persist across different ASR architectures, speech domains, or independent training runs. This makes it difficult to determine whether the method provides a broadly applicable pruning strategy or primarily benefits this particular setup.
  2. Insufficient isolation of the cumulative acoustic signal. Cumulative scores naturally depend on utterance duration as well as local acoustic variation. Although the paper compares against word-count and phone-count baselines, it does not directly establish how much of the improvement comes from acoustic dynamics rather than duration alone. This leaves the central explanation for the method's effectiveness partially unresolved.
Questions for authors
  1. Robustness: Can the authors report results across multiple independent ASR training seeds, including variance or confidence intervals, to establish whether the improvements over Teacher and Random selection are statistically robust?
  2. Duration confounding: Can the authors compare the proposed cumulative acoustic scores against a duration-only ranking baseline, or otherwise control for utterance duration, to isolate the contribution of acoustic variation?
  3. Generalization: Do the proposed criteria retain their advantages on another ASR architecture or speech dataset, particularly at the 5% retention budget where the paper reports its strongest gains?
Reviewer_yid3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses pruning ASR training data to a fixed utterance-count budget without the cost of model-based selection. It proposes three model-free deterministic acoustic criteria, accumulated spectral entropy, accumulated spectral flux, and log-Mel temporal variation, each mapping an utterance to a scalar via cumulative aggregation of frame-level statistics. Experiments with a Conformer on LibriSpeech show that at 5 and 1 percent retention these criteria outperform all baselines and stay competitive with a teacher-student criterion at larger budgets, while cutting selection time from over 13 GPU hours to under half an hour on CPU.

Critical strengths
  1. The method is simple, training-free, and GPU-free, and the paper directly quantifies the performance-efficiency trade-off, a practically relevant and honestly reported contribution.
  2. The counter-intuitive result that cheap signal statistics beat an expensive teacher criterion under aggressive pruning is interesting
Critical weaknesses
  1. The evidence is thin for a top venue, with a single dataset, a single architecture, and apparently single runs with no variance or significance testing, and the 1 percent regime with errors near 90 percent is nearly uninterpretable.
  2. The practical motivation is weakest where the method wins, because the win region at 1 to 5 percent retention yields error rates no practitioner would deploy, while at useful budgets the method is merely matched or beaten by the teacher. A cited reference also carries an impossible future arXiv identifier, which undermines confidence in the related-work rigor.
Questions for authors
  1. Do the findings transfer beyond read-speech LibriSpeech and the Conformer to a noisier or conversational corpus or a different model?
  2. Do the reported numbers come from single runs, and if you report mean and standard deviation over seeds does the best criterion retain a significant lead at 5 percent?
  3. Since the budget is an utterance count and the scores correlate with duration, can you show results controlling for total retained audio duration to separate more audio from more acoustic variation?
Reviewer_jQv9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a training-free, model-independent approach for training data selection for ASR. It rankstraining utterances using cumulative statistics computed directly from speech signals based on accumulated spectral entropy, spectral flux, and log-Mel temporal variation. Evaluated on LibriSpeech corpus using a Conformer architecture, the approach shows competitive performance with highly restrictive retention budgets.

Critical strengths
  1. The approach is well motivated by using cumulative rather than mean aggregation. It is also practical and efficient without requiring GPUs.
  2. Strong performance at low retention rates has been shown in the experiments.
Critical weaknesses
  1. The major concern is that the evaluation relies exclusively on LibriSpeech dataset, which a reading-style speech corpus recorded in relatively clean conditions. Whether the conclusions in the paper are still valid in more challenging scenarios (speech with background noises, conversational speech, domain mismatch between training and evaluation datasets).
  2. The downstream model is a specific Conformer architecture. Data selection can be model-dependent. It is not clear if the acoustic features are equally effective for other model architectures.
Questions for authors
  1. The approach is purely acoustic without linguistic content. At low retention rates, there might be a risk that the selected subsets of speech data lack of lexical diversity. Did you analyze the lexical coverage or phonemic balance of the subsets selected?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.