Grammatical Error Correction (GEC) and feedback play a vital role in supporting second language (L2) learners, educators, and examiners. While written GEC is well-established, spoken GEC (SGEC), aiming to provide feedback based on learners' speech, poses additional challenges due to disfluencies, transcription errors, and the lack of structured input. SGEC systems typically follow a cascaded pipeline consisting of Automatic Speech Recognition (ASR), disfluency detection, and GEC, making them vulnerable to error propagation across modules. This work examines an End-to-End (E2E) framework for SGEC and feedback generation, highlighting challenges and possible solutions when developing these systems. Cascaded, partial-cascaded and E2E architectures are compared, all built on the Whisper foundation model. A challenge for E2E systems is the scarcity of GEC labeled spoken data. To address this, an automatic pseudo-labeling framework is examined, increasing the training data from 77 to over 2500 hours. To improve the accuracy of the SGEC system, additional contextual information, exploiting the ASR output, is investigated. Candidate feedback of their mistakes is an essential step to improving performance. In E2E systems the SGEC output must be compared with an estimate of the fluent transcription to obtain the feedback. To improve the precision of this feedback, a novel reference alignment process is proposed that aims to remove hypothesised edits that results from fluent transcription errors. Finally, these approaches are combined with an edit confidence estimation approach, to exclude low-confidence edits. Experiments on the in-house Linguaskill (LNG) corpora and the publicly available Speak & Improve (S&I) corpus show that the proposed approaches significantly boost E2E SGEC performance.
Figures & tables
Figure 1: Illustration of the cascaded, partial-cascaded , and end-to-end SGEC systems.
Figure 2: Overview of the proposed end-to-end SGEC framework with SGEC and feedback evaluation.
Figure 3: The SGEC pseudo-labeling pipeline.
Figure 4: The end-to-end SGEC system prompting with additional ASR transcriptions. GEC reference: The cat calmly rested on the mat.
Figure 5: Illustration of the GEC reference modification process.
Corpus
Split
Hours
Speakers
Utts/Sents
Words
Spoken
LNG lbl
Train
77.6
1,908
34.8K
502K
Dev
7.8
176
3.3K
49K
Test
11.0
271
4.6K
69K
LNG unl V1
Train
2521.6
-
708,613
15M
LNG unl V2
Train
2585.6
-
716.9K
15M
S&I
Dev
20.8
-
6.9K
105K
Table 1: Statistics of datasets.
Model
small.en
large-v2
LNG
S&I
LNG
S&I
Whisper dsf + DD + GEC
13.45
17.35
12.11
14.49
Whisper flt + GEC
13.24
16.91
11.81
13.99
Whisper gec
13.48
17.76
11.10
13.21
Table 2: WER evaluation of cascaded, partial-cascaded , and end-to-end (E2E) SGEC models. All Whisper models are trained on LNG lbl .
Model
Approach
LNG
S&I
Whisper gec
Fine-tune
11.10
13.21
Qwen2-Audio-7B-Instruct
Zero-shot
26.10
-
LoRA
13.70
-
Voxtral-Small-24B-2507
Zero-shot
13.71
15.33
Table 3: WER evaluation of Whisper and multimodal LLMs.
Model
Training Set
small.en
large-v2
LNG
S&I
LNG
S&I
Whisper flt + GEC
13.24
16.91
11.81
13.99
Whisper gec
LNG lbl
13.48
17.76
11.10 †
13.21 †
LNG unl V1
14.16
18.11
12.93
15.92
+ LNG lbl
12.72
16.84
11.10 †
13.93
LNG unl V2
-
-
12.00
14.12
Table 4: Evaluation (WER) of Whisper gec performance with pseudo-labeled data on LNG lbl test and S&I dev sets. LNG unl V1 and V2 refer to the pseudo-labeled data generated based on the small.en and large-v2 models, respectively. † indicates a statistically significant improvement over the partial-cascaded system ( p<0.001 ).
LNG
S&I
LNG
S&I
Model Name
Prompt
WER
WER
P
R
F 0.5
P
R
F 0.5
Whisper flt + GEC
11.81
13.99
46.60
26.61
40.51
49.43
28.51
43.10
Whisper gec
✗
11.10 †
13.21 †
40.38
31.91
38.24
41.87
33.06
39.75
Whisper gec+pt
✓
11.08 †
13.09 †
43.92
32.63
41.08
45.56
33.88
42.62
Table 5: SGEC performance of E2E models with and without prompts, using Whisper large-v2. † denotes statistically significant improvement over the partial-cascaded system ( p<0.001 ).
Model
Train
LNG
S&I
Name
Ref
WER
P
R
F 0.5
WER
P
R
F 0.5
Whs flt + GEC
11.81
46.60
26.61
40.51
13.99
49.43
28.51
43.10
Whs gec
Orig
11.10
40.38
31.91
38.24
13.21
41.87
33.06
39.75
+ editavg
51.84
24.79
42.55
53.17
26.45
44.23
Whs gec+pt
Orig
11.08
43.92
32.63
41.08
13.09
45.56
33.88
42.62
+ editavg
51.10
27.58
43.65
54.08
28.20
45.69
Table 6: Feedback performance (Precision, Recall, F 0.5 ) of various end-to-end SGEC models using Whisper large-v2.
LNG
S&I
Setting
th
F 0.5
th
F 0.5
Dataset-specific
0.67
44.73
0.65
46.94
Cross-dataset
0.65
44.57
0.67
46.86
Table 7: Threshold robustness of edit-confidence filtering for the best-performing configuration (Whisper gec+pt with Strong reference alignment and edit avg ). The transferred threshold uses the optimal threshold from the other dataset.
Figure 6: Precision, Recall, and F 0.5 scores at different edit-confidence thresholds for the Whisper gec+pt model trained with Strong GEC references.
System
Output
Example 1
FLT Reference
eating habit are very important if you will have healthy
GEC Reference
eating habits are very important if you want to be healthy
Whs flt
is in habit are very important if you will have healthy
Whs flt + GEC
having a habit is very important if you want to healthy
Whs gec
eating habits are very important if you want to be healthy
Table 14
Model
Train
S&I
Name
Ref
WER
P
R
F 0.5
Whs flt + GEC
14.37
48.97
27.25
42.24
Whs gec+pt
Mild
13.44 †
47.76
29.96
42.69
Strong
13.72
53.61
26.29
44.39
+ editavg
57.35
24.50
45.22
Table 8: SGEC and Feedback performance on S&I Eval set. † denotes statistically significant improvement over the partial-cascaded system ( p<0.001 ).
Grammatical error correction (GEC) and explanation (GEE) have made rapid progress, but real teaching scenarios also require learner-friendly pedagogical feedback that is actionable, level-appropriate, and encouraging. We introduce SPFG (Spoken Pedagogical Feedback Generation), a dataset built based on the Speak & Improve Challenge 2025 corpus, pairing fluency-oriented transcriptions with GEC targets and human-verified teacher-style feedback, including preferred/rejected feedback pairs for preference learning. We study a transcript-based Spoken Grammatical Error Correction (SGEC) setting and evaluate three instruction-tuned LLMs (Qwen2.5, Llama-3.1, and GLM-4), comparing supervised fine-tuning (SFT) with preference-based alignment (using DPO and KTO) for jointly generating corrections and feedback. Results show that SFT provides the most consistent improvements, while DPO/KTO yield smaller or mixed gains, and that correction quality and feedback quality are weakly coupled. Our implementation is available at https://github.com/Skywalker-Harrison/spfg.
Junhong Liang, Yifan Lu, Ekaterina Kochmar +1
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE
Grammatical error correction (GEC) evaluation has traditionally relied on reference or edit overlap, which can penalize valid rewrites that differ from gold corrections. Reference-free metrics reduce this dependence, but evaluating whether a fluent output is a valid correction of the source remains challenging. We propose SURE, a source-conditioned reward evaluator trained on within-source preferences spanning minimal-edit and rewrite-oriented corrections. SURE jointly learns an overall reward with criteria-level supervision for grammaticality, faithfulness, and fluency, together with span-level grounding for source-side error resolution. Experiments on SEEDA show that SURE performs competitively against strong baselines, with particular gains on rewrite-style corrections and more disentangled criteria-level diagnostics. Our code is available at https://github.com/hayeonggg/SURE.
Hayeong Ryu, Sunhee Jo, Seunguk Yu +1
Department of Artificial Intelligence, Chung-Ang University · Graduate School of Advanced Imaging Sciences, Multimedia and Film, Chung-Ang University
A Grammatical Error Correction (GEC) system produces a sequence of edits to correct an erroneous sentence. The quality of these edits is typically evaluated against human annotations. However, a sentence may admit multiple valid corrections, and existing evaluation settings do not fully accommodate diverse application scenarios. Recent meta-evaluation approaches rely on human judgments across multiple references, but they are difficult to scale to large datasets. In this paper, we propose a new task, Scoring Edit Impact in GEC, which aims to automatically estimate the importance of edits produced by a GEC system. To address this task, we introduce a scoring framework based on an embedded association graph. The graph captures latent dependencies among edits and syntactically related edits, grouping them into coherent groups. We then perform perplexity-based scoring to estimate each edit's contribution to sentence fluency. Experiments across 4 GEC datasets, 4 languages, and 4 GEC systems demonstrate that our method consistently outperforms a range of baselines. Further analysis shows that the embedded association graph effectively captures cross-linguistic structural dependencies among edits.
Qiyuan Xiao, Xiaoman Wang, Yunshi Lan
School of Data Science and Engineering, East China Normal University