EdgeSpot: Efficient and High-Performance Few-Shot Model for Keyword Spotting

Alican Gok; Oguzhan Buyuksolak; Osman Erman Okman

REVIEW 3 major objections 3 minor 22 references

EdgeSpot: Efficient and High-Performance Few-Shot Model for Keyword Spotting

T0 review · 3 major / 3 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read EdgeSpot lifts few-shot keyword spotting accuracy at low false-alarm rates without heavy compute.

desk verdict Solid incremental gain for few-shot KWS on edge devices; the Fused BC-ResBlock plus trainable PCEN and temporal attention is genuinely new, but the baseline tuning parity and missing PCEN ablation need to be addressed before trusting the headline 73.7->82.0. read the letter →

arxiv 2601.16316 v1 pith:DNVSH3G3 submitted 2026-01-22 eess.AS cs.CLcs.SD

Oguzhan Buyuksolak , Alican Gok , Osman Erman Okman This is my paper · ORCID

classification eess.AScs.CLcs.SD

keywords few-shotkeywordspottingedgecomputingknowledgedistillationself-supervisedspeechPCENtemporalself-attentionBC-ResNetfalse-alarmrate

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that a small, edge-oriented neural network can get substantially better few-shot keyword spotting accuracy at low false-alarm rates than a strong, widely used baseline, without adding heavy compute. The proposed model, EdgeSpot, keeps a BC-ResNet-style acoustic backbone and adds three targeted changes: a trainable Per-Channel Energy Normalization (PCEN) frontend, fused early residual blocks, and a lightweight temporal self-attention head. Trained by knowledge distillation from a self-supervised teacher with Sub-center ArcFace, EdgeSpot consistently beats re-trained BC-ResNet models on both in-domain and cross-domain tests. The largest variant, EdgeSpot-4, raises 10-shot accuracy at 1% false alarms from 73.7% to 82.0% on a cross-domain test set, while running at 29.4M MACs with 128k parameters. If correct, this gives edge devices a way to enroll new keywords from a few examples without surrendering low false-alarm behavior.

What carries the argument

The argument rests on three architectural modifications applied to the BC-ResNet backbone. First, a trainable Per-Channel Energy Normalization (PCEN) frontend replaces static log-mel compression, giving per-channel automatic gain control and root compression that helps transfer across domains. Second, early BC-ResBlocks are 'fused': the depthwise temporal convolution plus 1x1 pointwise projection is replaced by a single regular temporal convolution, an early-fusion idea that simplifies optimization at negligible cost. Third, a lightweight temporal scaled dot-product attention head, fed by a depthwise 1D convolutional relative positional encoding, encodes long-range temporal dependencies and

What would settle it

Run the same few-shot keyword spotting protocol but give the BC-ResNet baselines a per-model hyperparameter search (learning rate, SpecAugment mask sizes, number of epochs) comparable to the effort put into EdgeSpot. If the 73.7% to 82.0% gap at 1% FAR shrinks to near zero, the paper's central claim about architectural advantage is falsified. Alternatively, ablate the attention head from EdgeSpot and re-test: if removing it leaves low-FAR accuracy basically unchanged, the attention head is not doing the work attributed to it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that pairing an otherwise unchanged BC-ResNet backbone with a trainable PCEN frontend, a fused early-block design, and a single-head temporal self-attention with convolutional relative positional encoding yields consistently better few-shot keyword spotting accuracy at a fixed false-alarm rate than the BC-ResNet baseline itself, across all four model widths and on both the in-domain Multilingual Spoken Words Corpus and the cross-domain Google Speech Commands set. The improvement is largest in the strict low-FAR regime: EdgeSpot-4 improves the 10-shot accuracy at 1% FAR from 73.7% to 82.0% on GSC while using only 29.4M MACs and 128k parameters. Under

Load-bearing premise

The chief load-bearing premise is that the BC-ResNet baselines were trained fairly: the authors re-trained them with the same loss, data splits, and hyperparameter schedule as EdgeSpot, but give no evidence that the baselines received equivalent tuning effort, so an undertuned baseline would inflate the reported accuracy gap.

Editorial extensions

If this is right

If the reported gains hold, low-FAR accuracy in few-shot keyword spotting can be improved without a large compute penalty; EdgeSpot-4 runs at 29.4M MACs and 128k parameters.
Cross-domain robustness improves: EdgeSpot beats BC-ResNet on the Google Speech Commands set even though it is trained on MSWC, suggesting the frontend and attention help generalize to new recording conditions.
At relaxed operating points and with 10-shot enrollment, EdgeSpot-4 reaches or exceeds the teacher model's accuracy, meaning a tiny student can recover most of the discriminative power of a large self-supervised encoder.
The improvements are consistent across model widths (EdgeSpot-1 to -4), so the recipe scales down to even smaller footprints if needed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The PCEN frontend is architecture-agnostic: adding it to other acoustic backbones might yield similar cross-domain gains, especially in noisy conditions where PCEN is known to help.
The fused early-block change suggests that the standard depthwise-separable stack in the earliest stages is not just redundant but slightly harmful for optimization; this could transfer to other residual speech models.
The temporal self-attention head may become more valuable for longer utterances or continuous speech, where local convolutions miss long-range dependencies, an extension not tested in this 1-second KWS setting.
A stronger test would be to evaluate EdgeSpot under noisy or far-field conditions, where PCEN's gain control should be most visible; the paper does not report such conditions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a genuine, incremental improvement for few-shot KWS on edge devices, not a breakthrough. The new bits are the Fused BC-ResBlock (EfficientNetV2-style early fusion applied to the BC-ResNet temporal branch) and the specific stack of trainable PCEN, depthwise positional encoding, and single-head temporal SDPA. On both MSWC and GSC, every EdgeSpot variant beats the retrained BC-ResNet baseline at fixed FAR, and the headline 10-shot GSC ACC@1% gain (73.7 to 82.0 at 29.4M MACs) exceeds the reported std. That is a real result if the controls are honest.

What the paper does well: it is clearly written, the ablation logic is sensible, and the evaluation protocol is consistent with their prior work [6]. They retrained BC-ResNet under the same teacher-student KD setup rather than copying published numbers, and they report both DET on MSWC and ACC/FAR on cross-domain GSC. The cost increase over BC-ResNet is small (e.g., 24.1M vs 29.4M MACs for tau=4), which is the right trade to claim.

Soft spots, in order of importance. First, the baseline tuning parity is asserted, not shown. They say 'for fairness' they adapted BC-ResNet, but there is no hyperparameter search or evidence that the same LR, SpecAugment, and lambda are equally good for both architectures. Shared settings can favor the model with the fused early blocks. If BC-ResNet-4 was undertuned, the 73.7->82.0 gap is inflated. This is a legitimate concern and the most load-bearing condition for the headline. It is not fatal: the gains are consistent across all four scales and both datasets, including 1-shot, which would be an odd pattern for pure tuning artifacts. Second, the PCEN-only ablation is mentioned in Section 2.1.1 but never shown. That is easy to fix. Third, the final embedding head is under-specified; Table 1 says the final conv1d has output channel 1, which contradicts the 64-D embedding described in the text. Fourth, no code or checkpoints, so re-implementation depends on [6].

The 'nearly matches Teacher' discussion is partly by construction, since the student is trained to minimize embedding distance to that teacher, so do not count it as independent evidence.

Verdict: this deserves a serious referee. The central claim is empirical, internally consistent, and not forced by any equation. The required fixes are verification tasks, not conceptual ones: show the PCEN ablation, document the baseline tuning effort, and ideally release code. I would not desk-reject it.

Referee Report

3 major / 3 minor

Summary. The paper introduces EdgeSpot, a few-shot keyword spotting (FS-KWS) model that combines a BC-ResNet acoustic backbone with three modifications: a trainable Per-Channel Energy Normalization (PCEN) frontend, fused early residual blocks, and a lightweight temporal self-attention head with convolutional relative positional encoding. Models are trained via knowledge distillation from a Wav2Vec2-based teacher with Sub-center ArcFace loss and evaluated under a prototype-based 1-/10-shot protocol on MSWC (in-domain) and Google Speech Commands (cross-domain). The central claim is that EdgeSpot consistently outperforms retrained BC-ResNet baselines, with the headline result that EdgeSpot-4 improves GSC 10-shot accuracy at 1% FAR from 73.7% to 82.0% at 29.4M MACs and 128k parameters.

Significance. If the reported results are reproducible, the paper provides a practically useful recipe for improving low-FAR accuracy in few-shot KWS on edge devices, at a modest computational overhead (e.g., +5.3M MACs and +31.7k parameters for EdgeSpot-4 vs BC-ResNet-4). The evaluation is reasonably thorough: multiple model scales, cross-domain testing, and repeated random trials with reported standard deviations. The headline gain substantially exceeds the reported variability. However, the central empirical claim depends on the fairness of the retrained BC-ResNet controls, and one of the three proposed components is asserted to help but never shown in an ablation. These issues need to be resolved before the comparative claim is fully convincing.

major comments (3)

[Section 3, Table 2] The fairness of the BC-ResNet baseline comparison is the load-bearing premise of the paper. All models share one learning-rate schedule, one distillation weight λ=5e-5, and width-tied SpecAugment, but no evidence is given that the BC-ResNet baselines received comparable hyperparameter tuning. If BC-ResNet-4 is undertuned, the reported 73.7→82.0 gain at GSC 10-shot ACC@1% is inflated. Please provide either (a) a per-architecture hyperparameter search (e.g., over LR and λ) with the best result for each model, or (b) an external calibration showing the retrained baselines are not weaker than known FS-KWS results (e.g., comparing to [3,4] or the original BC-ResNet paper's protocol).
[Section 2.1.1] The text asserts: 'In initial ablations, adding only PCEN with no other architectural change improved cross-domain test set performance (GSC, see section 2.2).' No such ablation is shown anywhere, and the reference to section 2.2 points to training details, not results. Since PCEN is one of the three proposed modifications, the paper should include a decomposition ablation (backbone with 0, 1, 2, or 3 of the proposed additions) to support this claim and to attribute the reported gains. Without it, the statement is unverifiable.
[Section 2.1.4 / Table 1] The architecture description is internally inconsistent regarding the final embedding. Section 2.1.4 states that the head produces an utterance-level 64-dimensional embedding e∈R^64, and Fig. 1 labels the final block '64 Dim. Embedding'. However, Table 1 and its footnote state that the final conv1d layer always has an output channel of 1 (the c column for that row is 1). A conv1d with 1 output channel would not yield a 64-D embedding as written. Please clarify the exact final operations (e.g., a 64-channel conv1d followed by global pooling, or a different aggregation scheme) and correct the table/footnote accordingly.

minor comments (3)

[Abstract / Section 3] The abstract claims EdgeSpot 'consistently' outperforms BC-ResNet at fixed FAR. In Table 2, a few cells are ties (e.g., MSWC DET5% 10-shot for τ=4: 99.1 vs 99.1) or slightly worse (e.g., GSC AUC 1-shot for τ=3: 73.2 vs 73.7). Consider softening 'consistently' to 'in most settings' or restricting the claim to the low-FAR metrics emphasized in the paper.
[Table 1 / Notation] The table lists 'Edgespot' with lowercase 's' in several rows (e.g., 'Edgespot-1', 'Edgespot-4') while the text uses 'EdgeSpot'. Please unify capitalization.
[Section 2.2] SpecAugment is applied with 'settings tied to model width' (no augmentation for τ=1, and fixed F=6, T=8 for τ≥2). The paper does not explain why the augmentation should depend on width; a brief justification or ablation would help the reader assess the fairness of sharing these settings across architectures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EdgeSpot's reported gains are an empirical comparison against re-trained BC-ResNet baselines under an identical distillation loss; no equation or fitted parameter forces the outcome.

full rationale

The paper's central claim—that EdgeSpot improves low-FAR accuracy over BC-ResNet—is established by measured results in Table 2. The student is trained with Eq. (9), L = L_KD + lambda*L_SCAF, but this same loss is applied to EdgeSpot and to the BC-ResNet baselines ('For fairness, we adapt BC-ResNet to the FS-KWS setting by replacing the final classifier to produce 64-dimensional embeddings compatible with our KD setup.'). Thus any performance difference must come from the actual forward architectures (trainable PCEN, fused early BC-ResBlocks, temporal SDPA with positional encoding), not from the training objective itself. The statements that EdgeSpot 'nearly matches the Teacher' are an expected consequence of the KD term in Eq. (9), but they are presented as empirical observations in the Discussion and are not the headline claim; the headline is an EdgeSpot-versus-BC-ResNet comparison, not a derivation of teacher matching. The use of the authors' prior work [6] for the teacher and evaluation protocol is a normal methodological dependency, not a circular argument: the results are measured on external benchmarks (MSWC and GSC), and the reference is not invoked to prove the architectural superiority. Concerns about possible under-tuning of the BC-ResNet baselines are a threat to experimental validity, not an instance of circularity, because no equation or fitted parameter is shown to reproduce the reported 73.7-to-82.0 gain by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The headline numbers rest on assumptions imported from prior work: the teacher from [6] provides good distillation targets; the prototype/threshold protocol from [6] is a valid few-shot decision rule; BC-ResNet [10], PCEN [12], and SCAF [17] behave as described in their source papers; and the MSWC/GSC splits are used as-is. The only hand-picked numbers that affect the result are training/architecture hyperparameters (lambda=5e-5, d=64, kappa=16, SpecAugment settings, PCEN learned scalars); none is fitted to reproduce the headline 82.0% number. No new entities are invented.

free parameters (4)

distillation weight lambda = 5e-5 (chosen by hand)
Balances KD and SCAF losses in Eq. (9); no search reported; central to the training objective.
trainable PCEN scalars (alpha, r, delta) and smoothing s = learned end-to-end, channel-shared
PCEN parameters are trained jointly with the model (Section 2.1.1); any effect of PCEN depends on these learned values.
attention projection dimension d and positional kernel kappa = d=64, kappa=16 (padding 8)
Chosen by hand (Sections 2.1.3-2.1.4); set the capacity of the attention head.
SpecAugment settings for tau>=2 = F=6, T=8, time-stretch 0.9-1.1
Stated in Section 2.2; augmentation strength is tied to model width; applies to all compared models but is a free choice of the authors.

assumptions (5)

domain assumption Teacher model from [6] (Wav2Vec2.0 through layer 16 + SCAF dimensionality-reduction head) produces embeddings worth distilling.
The whole student pipeline minimizes distance to these embeddings (Section 2.2); if the teacher is weak or mis-calibrated, every student number moves with it. Teacher evaluation is reused from [6], not re-established here.
domain assumption Prototype averaging of K enrollment embeddings plus distance threshold at a chosen FAR is a valid few-shot decision rule.
Inherited from [6] and applied identically to all models (Sections 2.2, 3); standard in the FS-KWS literature but an assumption nonetheless.
domain assumption 40-band mel-spectrogram (40x101) computed as described preserves the information needed to match teacher embeddings.
Student input differs from the teacher's waveform input; the distillation must bridge that domain gap, which is assumed learnable (Section 2.2).
domain assumption Sub-center ArcFace [17] with distillation behaves as published and its default hyperparameters transfer to this setting.
SCAF is applied without modification (Section 2.2); any implementation or hyperparameter mismatch would affect both EdgeSpot and baselines equally but is untested here.
domain assumption MSWC English train/test and GSC splits used as-is provide a fair few-shot evaluation with no unintended overlap.
Models are trained on MSWC and tested on GSC for cross-domain (Section 2.2); dataset curation assumptions are taken from the dataset papers.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EdgeSpot: Efficient and High-Performance Few-Shot Model for Keyword Spotting." pith.science (2026). https://pith.science/paper/DNVSH3G3

@misc{pith2026260116316,
  author       = {Pith},
  title        = {Pith review of: EdgeSpot: Efficient and High-Performance Few-Shot Model for Keyword Spotting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DNVSH3G3}},
  note         = {Machine review of arXiv:2601.16316}
}

read the original abstract

We introduce an efficient few-shot keyword spotting model for edge devices, EdgeSpot, that pairs an optimized version of a BC-ResNet-based acoustic backbone with a trainable Per-Channel Energy Normalization frontend and lightweight temporal self-attention. Knowledge distillation is utilized during training by employing a self-supervised teacher model, optimized with Sub-center ArcFace loss. This study demonstrates that the EdgeSpot model consistently provides better accuracy at a fixed false-alarm rate (FAR) than strong BC-ResNet baselines. The largest variant, EdgeSpot-4, improves the 10-shot accuracy at 1% FAR from 73.7% to 82.0%, which requires only 29.4M MACs with 128k parameters.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 5 linked inside Pith

[6]

Fully unsupervised training of few-shot keyword spotting,

Dongjune Lee, Minchan Kim, Sung Hwan Mun, Min Hyun Han, and Nam Soo Kim, “Fully unsupervised training of few-shot keyword spotting,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 266–272

2023
[1]

Adapting the model for different sets of keywords requires repeating the same training procedure; therefore, it is not feasible to adapt the models on edge hardware

INTRODUCTION Keyword Spotting (KWS) enables hands-free interaction by detecting specific spoken commands or wake-words in audio, but conventional neural network (NN) based systems require extensive data and significant computational resources during training. Adapting the model for different sets of keywords requires repeating the same training procedure;...

2026
[2]

PROPOSED METHOD 2.1. Model Architecture EdgeSpot adopts BC-ResNet as its acoustic backbone and in- troduces three targeted, edge-friendly improvements that raise arXiv:2601.16316v1 [eess.AS] 22 Jan 2026 PCEN 5x5 Conv2D 40x101 Mel-Spec. Fused BC-Res. BC-Res. 5x5 DW Conv2D 1x1 Conv2D DW Conv1D Pos. Enc. Q K V SDPA Conv1D 64 Dim. Embedding PReLU BN ReLUBN Re...

arXiv 2026
[3]

For fairness, we adapt BC-ResNet to the FS- KWS setting by replacing the final classifier to produce 64- dimensional embeddings compatible with our KD setup

EXPERIMENTAL RESULTS We evaluate EdgeSpot against BC-ResNet baselines under the same prototype-based FS-KWS protocol as our prior work [6]. For fairness, we adapt BC-ResNet to the FS- KWS setting by replacing the final classifier to produce 64- dimensional embeddings compatible with our KD setup. Table 2 summarizes 1-shot and 10-shot results on MSWC and G...
[4]

EdgeSpot delivers higher low-FAR accuracy than BC-ResNet on both MSWC and cross-domain GSC with minimal additional on-device cost

DISCUSSION We evaluated EdgeSpot under the same prototype-based FS-KWS protocol as our prior work [6] and, for fairness, re-trained BC-ResNet baselines with the identical KD+SCAF loss tailored to FS-KWS. EdgeSpot delivers higher low-FAR accuracy than BC-ResNet on both MSWC and cross-domain GSC with minimal additional on-device cost. Notably, EdgeSpot-4 ou...
[5]

Few-shot keyword spotting in any language,

Mark Mazumder, Colby Banbury, Josh Meyer, Pete Warden, and Vijay Janapa Reddi, “Few-shot keyword spotting in any language,” inInterspeech, 2021, pp. 4214–4218

2021
[7]

On-device cus- tomization of tiny deep learning models for keyword spotting with few examples,

Manuele Rusci and Tinne Tuytelaars, “On-device cus- tomization of tiny deep learning models for keyword spotting with few examples,”IEEE Micro, vol. 43, no. 6, pp. 50–57, 2023

2023
[8]

Few-shot open- set learning for on-device customization of keyword spotting systems,

Manuele Rusci and Tinne Tuytelaars, “Few-shot open- set learning for on-device customization of keyword spotting systems,” inProc. Interspeech, 2023, pp. 2768– 2772

2023

Show all 22 references

[9]

Deep residual learning for image recognition,

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” CoRR, vol. abs/1512.03385, 2015

2015 arXiv
[10]

Enhancing few- shot keyword spotting performance through pre-trained self-supervised speech models,

Alican Gok, Oguzhan Buyuksolak, Osman Erman Okman, and Murat Saraclar, “Enhancing few- shot keyword spotting performance through pre-trained self-supervised speech models,”arXiv preprint arXiv:2506.17686, 2025

2025
[11]

Match- boxNet: 1D time-channel separable convolutional neu- ral network architecture for speech commands recogni- tion,

Somshubra Majumdar and Boris Ginsburg, “Match- boxNet: 1D time-channel separable convolutional neu- ral network architecture for speech commands recogni- tion,” inProc. Interspeech, 2020, pp. 3356–3360

2020
[12]

Depthwise sepa- rable convolutional resnet with squeeze-and-excitation blocks for small-footprint keyword spotting,

Menglong Xu and Xiao-Lei Zhang, “Depthwise sepa- rable convolutional resnet with squeeze-and-excitation blocks for small-footprint keyword spotting,” inInter- speech 2020, 2020, pp. 2547–2551

2020
[13]

Temporal convolution for real- time keyword spotting on mobile devices,

Seungwoo Choi, Seokjun Seo, Beomjun Shin, Hyeong- min Byun, Martin Kersner, Beomsu Kim, Dongyoung Kim, and Sungjoo Ha, “Temporal convolution for real- time keyword spotting on mobile devices,” inProc. In- terspeech, 2019, pp. 3372–3376

2019
[14]

Broadcasted residual learning for ef- ficient keyword spotting,

Byeonggeun Kim, Simyung Chang, Jinkyu Lee, and Dooyong Sung, “Broadcasted residual learning for ef- ficient keyword spotting,” inProc. Interspeech 2021, 2021, pp. 4538–4542

2021
[15]

Speech commands: A dataset for limited-vocabulary speech recognition,

Pete Warden, “Speech commands: A dataset for limited-vocabulary speech recognition,”arXiv preprint arXiv:1804.03209, 2018

2018 arXiv
[16]

Trainable frontend for robust and far-field keyword spotting,

Yuxuan Wang, Pascal Getreuer, Thad Hughes, Richard F. Lyon, and Rif A. Saurous, “Trainable frontend for robust and far-field keyword spotting,” CoRR, vol. abs/1607.05666, 2016

2016 arXiv
[17]

Multilingual spoken words corpus,

Mark Mazumder, Sharad Chitlangia, Colby Banbury, Yiping Kang, Juan Manuel Ciro, Keith Achorn, Daniel Galvez, Mark Sabini, Peter Mattson, David Kanter, et al., “Multilingual spoken words corpus,” inThirty- fifth Conference on Neural Information Processing Sys- tems Datasets and...

2021
[18]

Per-channel energy normalization: Why and how,

Vincent Lostanlen, Justin Salamon, Mark Cartwright, Brian McFee, Andrew Farnsworth, Steve Kelling, and Juan Pablo Bello, “Per-channel energy normalization: Why and how,”IEEE Signal Processing Letters, vol. 26, no. 1, pp. 39–43, 2019

2019
[19]

Efficientnetv2: Smaller models and faster training,

Mingxing Tan and Quoc Le, “Efficientnetv2: Smaller models and faster training,” inInternational conference on machine learning. PMLR, 2021, pp. 10096–10106

2021
[20]

wav2vec 2.0: A framework for self- supervised learning of speech representations,

Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli, “wav2vec 2.0: A framework for self- supervised learning of speech representations,” inAd- vances in Neural Information Processing Systems, 2020, vol. 33, pp. 12449–12460

2020
[21]

Sub-center arcface: Boosting face recognition by large-scale noisy web faces,

Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou, “Sub-center arcface: Boosting face recognition by large-scale noisy web faces,” inEuropean Conference on Computer Vision (ECCV), 2020, pp. 741–757

2020
[22]

Specaugment: A simple data augmentation method for automatic speech recognition,

Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le, “Specaugment: A simple data augmentation method for automatic speech recognition,”arXiv preprint arXiv:1904.08779, 2019

1904 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [6]

Fully unsupervised training of few-shot keyword spotting,

Dongjune Lee, Minchan Kim, Sung Hwan Mun, Min Hyun Han, and Nam Soo Kim, “Fully unsupervised training of few-shot keyword spotting,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 266–272

2023

[2] [1]

Adapting the model for different sets of keywords requires repeating the same training procedure; therefore, it is not feasible to adapt the models on edge hardware

INTRODUCTION Keyword Spotting (KWS) enables hands-free interaction by detecting specific spoken commands or wake-words in audio, but conventional neural network (NN) based systems require extensive data and significant computational resources during training. Adapting the model for different sets of keywords requires repeating the same training procedure;...

2026

[3] [2]

PROPOSED METHOD 2.1. Model Architecture EdgeSpot adopts BC-ResNet as its acoustic backbone and in- troduces three targeted, edge-friendly improvements that raise arXiv:2601.16316v1 [eess.AS] 22 Jan 2026 PCEN 5x5 Conv2D 40x101 Mel-Spec. Fused BC-Res. BC-Res. 5x5 DW Conv2D 1x1 Conv2D DW Conv1D Pos. Enc. Q K V SDPA Conv1D 64 Dim. Embedding PReLU BN ReLUBN Re...

arXiv 2026

[4] [3]

For fairness, we adapt BC-ResNet to the FS- KWS setting by replacing the final classifier to produce 64- dimensional embeddings compatible with our KD setup

EXPERIMENTAL RESULTS We evaluate EdgeSpot against BC-ResNet baselines under the same prototype-based FS-KWS protocol as our prior work [6]. For fairness, we adapt BC-ResNet to the FS- KWS setting by replacing the final classifier to produce 64- dimensional embeddings compatible with our KD setup. Table 2 summarizes 1-shot and 10-shot results on MSWC and G...

[5] [4]

EdgeSpot delivers higher low-FAR accuracy than BC-ResNet on both MSWC and cross-domain GSC with minimal additional on-device cost

DISCUSSION We evaluated EdgeSpot under the same prototype-based FS-KWS protocol as our prior work [6] and, for fairness, re-trained BC-ResNet baselines with the identical KD+SCAF loss tailored to FS-KWS. EdgeSpot delivers higher low-FAR accuracy than BC-ResNet on both MSWC and cross-domain GSC with minimal additional on-device cost. Notably, EdgeSpot-4 ou...

[6] [5]

Few-shot keyword spotting in any language,

Mark Mazumder, Colby Banbury, Josh Meyer, Pete Warden, and Vijay Janapa Reddi, “Few-shot keyword spotting in any language,” inInterspeech, 2021, pp. 4214–4218

2021

[7] [7]

On-device cus- tomization of tiny deep learning models for keyword spotting with few examples,

Manuele Rusci and Tinne Tuytelaars, “On-device cus- tomization of tiny deep learning models for keyword spotting with few examples,”IEEE Micro, vol. 43, no. 6, pp. 50–57, 2023

2023

[8] [8]

Few-shot open- set learning for on-device customization of keyword spotting systems,

Manuele Rusci and Tinne Tuytelaars, “Few-shot open- set learning for on-device customization of keyword spotting systems,” inProc. Interspeech, 2023, pp. 2768– 2772

2023