Noise-Robust Speech Transcription with Quantized Language Model

Correction for Industrial Settings

Marco Murgia

1,2

, Marco Fontana

, Alberto Pes

, Diego Reforgiato Recupero

1,2

, Giuseppe Scarpi

1,2

and Leonardo Daniele Scintilla

Department of Mathematics and Computer Science, University of Cagliari, Via Ospedale 72, Cagliari, Italy

R2M Solution S.r.l., Via Fratelli Cuzio, 42, Pavia, Italy

Fontana Group, Via Alcide de Gasperi 16, Calolziocorte, Italy

Keywords:

Language Models, Automatic Speech Recognition, Synthetic Dataset, Noisy Environment.

Abstract:

In this paper, we propose a robust and computationally efﬁcient pipeline for transcribing speech in noisy en-

vironments, such as workshops and industrial settings. The pipeline is designed to operate ofﬂine, making

it suitable for resource-constrained scenarios. It begins with a noise ﬁltering module that preprocesses audio

recordings to suppress background noise and enhance speech clarity. The ﬁltered audio is then passed to an

Automatic Speech Recognition (ASR) model, which generates initial transcription outputs. Given the potential

for transcription errors in challenging acoustic conditions, we incorporate a quantized Small Language Model

(SLM) trained on an ontology of defects related to the industrial environment to post-process and correct

these errors. The quantization of the SLM signiﬁcantly reduces its computational footprint while maintain-

ing correction accuracy, enabling the pipeline to function effectively on low-resource devices. Experimental

evaluations demonstrate the effectiveness of the proposed approach in improving transcription quality in noisy

conditions, highlighting its practicality for ofﬂine and resource-limited applications. In fact, preliminary vali-

dation on a synthetic dataset of 200 sentences in Italian and English showed a consistent F1 score of 87.04%

for SNR as challenging as -5 dBW (Decibels Watt) in Italian sentences and 91.25% in English sentences, with

the least computationally expensive version of Whisper (Whisper Tiny) and the SLM correction.

1 INTRODUCTION

Automatic Speech Recognition (ASR) systems of-

ten struggle in noisy industrial environments like

workshops, where machinery sounds and reverbera-

tions lead to frequent transcription errors (Li et al.,

2014; Virtanen et al., 2017; Mehrish et al., 2023).

Furthermore, many state-of-the-art ASR systems

rely on cloud processing, which is often unfeasi-

ble for scenarios with limited hardware resources

or privacy-sensitive ones that require ofﬂine opera-

tion (Bodepudi et al., 2019). While noise suppression

and robust models have improved (Zhang et al., 2018;

Ephraim and Malah, 1984), a comprehensive solution

for accurate, ofﬂine transcription remains a critical

challenge for industrial applications in quality control

and safety compliance (Huang et al., 2014). To ad-

dress these challenges, we propose a robust and com-

putationally efﬁcient pipeline for transcribing speech

commands in noisy settings. In collaboration with

Fontana Group, an Italian leader in automotive manu-

facturing, we developed a toolchain to help operators

recognize car body defects. The main challenges are

the noisy workshop environment and the need for a

solution that works locally on systems with limited

hardware, given the potential lack of a reliable inter-

net connection. Our pipeline integrates three compo-

nents: (i) a noise ﬁltering module, (ii) an ASR model

for the transcription, and (iii) a quantized Small Lan-

guage Model (SLM) trained on a domain-speciﬁc de-

fect ontology to correct errors. This quantized SLM

ensures a reduced computational footprint, enabling

effective on-device performance. Preliminary valida-

tion on a synthetic test set demonstrates the effective-

ness of our approach, achieving F1-scores of 87.04%

for Italian and 91.25% for English sentences in sce-

narios with an SNR as low as -5 dBW.

Murgia, M., Fontana, M., Pes, A., Recupero, D. R., Scarpi, G. and Scintilla, L. D.

Noise-Robust Speech Transcription with Quantized Language Model Correction for Industrial Settings.

DOI: 10.5220/0013704200003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 2, pages 479-486

ISBN: 978-989-758-770-2; ISSN: 2184-2809

479

2 RELATED WORK

Robust ASR in noisy industrial settings (Orel and

Varol, 2023; Dua et al., 2023; Bandela et al.,

2023) has been addressed through various techniques.

Front-end approaches, ranging from traditional spec-

tral subtraction (Moore et al., 2017) and Wiener ﬁl-

tering (Gomez and Kawahara, 2010) to modern deep

learning denoisers (G. et al., 2024), aim to enhance

speech intelligibility but can be computationally de-

manding for ofﬂine, resource-constrained contexts.

On the ASR side, while large models like the Whis-

per family (Radford et al., 2022) offer impressive

robustness, their computational cost remains a bar-

rier for edge deployment, and lightweight variants

like Whisper Tiny

often lack accuracy in challeng-

ing conditions, especially with domain-speciﬁc ter-

minology. To bridge this gap, post-processing with

language models (LLMs/SLMs) has emerged as a

powerful error correction strategy (Ma et al., 2023;

Yang et al., 2023; Jiang and Poellabauer, 2021), with

domain-speciﬁc models showing particular promise

for adapting to specialized vocabularies (Suh et al.,

2024). Furthermore, optimizing SLMs via quanti-

zation techniques (Andreyev, 2025; Gholami et al.,

2022) is critical for deployment on low-power de-

vices. However, prior work has rarely focused on

the combined application of these components: noise

ﬁltering, lightweight ASR, and quantized, domain-

speciﬁc SLM correction within a ofﬂine pipeline (Rao

et al., 2020; Kamahori et al., 2025). Our work bridges

this gap by proposing and validating such a pipeline

tailored for noise-robust, ofﬂine industrial applica-

tions.

3 BACKGROUND

3.1 The Targeted Task

Let S be the space of voice signals acquired during car

body inspection, where each signal s ∈ S is modeled

as:

s = v + n (1)

with v a spoken phrase containing a relevant term, and

n background noise.

The goal is to ﬁnd the correct transcription

t of the

term in v given an observed signal s via a function:

t = f (s) (2)

where f : S → T is a speech recognition and correc-

tion function, and T represents a ﬁnite set of domain-

speciﬁc terms belonging to a closed vocabulary.

https://huggingface.co/openai/whisper-tiny

Each phrase includes exactly one term from T , in

either Italian or English:

• English: positive bulge, negative bulge, wavi-

ness, deformation, reworking traces, crack, fail-

ure, scratches, glue residues, orange peel effect

• Italian: bollo positivo, bollo negativo, ondu-

lazione, deformazione, tracce, buco, rottura, graf-

ﬁo, residuo di colla, effetto buccia d’arancia

This is a single-label, multi-class classiﬁcation

task, where each input must be assigned to one class

from the corresponding dictionary.

3.2 Material

Our pipeline leverages three key models, speciﬁcally

optimized for an efﬁcient workﬂow. For noise sup-

pression, we employ DeepFilterNet

(Schröter et al.,

2022), a speech enhancement framework. Initial tran-

scription is performed by Whisper

, using the op-

timized Faster-Whisper

implementation integrated

with WhisperX (Bain et al., 2023) for fast inference

and accurate alignment. For the ﬁnal correction step,

we use the lightweight SLM Gemma 2 2B-it

. To

ensure its suitability for on-device deployment, we

quantized the model to a 4-bit GGUF format

and ran

it using Llama.cpp

with its Python bindings

. This

setup signiﬁcantly reduces the model’s memory foot-

print, making it suitable for on-device inference.

4 THE PROPOSED PIPELINE

The pipeline operates in three steps. First, incom-

ing audio, potentially containing background noise, is

processed by a Denoiser Module based on DeepFilter-

Net to enhance speech clarity. Second, the ﬁltered au-

dio is passed to an ASR Module, which uses Whisper

to generate an initial text transcription. Finally, this

transcription is fed into a Corrector Module, which

leverages a ﬁne-tuned Gemma SLM to identify and

extract the correct defect term from our predeﬁned vo-

cabulary, reﬁning the ﬁnal output.

https://github.com/Rikorose/DeepFilterNet

https://github.com/openai/whisper

https://github.com/SYSTRAN/faster-whisper

https://huggingface.co/google/gemma-2-2b-it

https://huggingface.co/docs/hub/en/gguf

https://github.com/ggml-org/llama.cpp

https://github.com/abetlen/llama-cpp-python

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

480

5 DATASET

5.1 Training Set

To ﬁne-tune the Gemma SLM as a corrector, we con-

structed a synthetic dataset to map potentially erro-

neous transcriptions to our domain-speciﬁc vocabu-

lary. The process began with the generation of a set

of clean base sentences using the Claude-Sonnet 3.5

LLM

. For each of the 10 English and 10 Italian de-

fect terms (see Section 3), we prompted the LLM

to create 30 unique example sentences relevant to a

car body inspection context (e.g., "I see a deforma-

tion in the hood of the car"), ensuring each sentence

contained only one vocabulary term. This resulted in

600 unique base sentences (300 per language), which

then served as the foundation for introducing pro-

grammatic perturbations.

Programmatic Perturbation for Realism

To ensure the SLM corrector is robust to real-world

ASR errors, its training data have to reﬂect such im-

perfections. Therefore, we programmatically per-

turbed our 600 clean base sentences to simulate typi-

cal transcription errors. With a custom Python script,

we introduced a set of perturbations, including char-

acter substitutions (e.g., deformation -> defomation),

character deletions (e.g., scratches -> scrtches), and

incorrect word splitting/merging. For each clean sen-

tence, we generated 12 perturbed variants with a de-

gree of distortion, from single minor alterations to

multiple, disruptive errors. This process yielded a ﬁ-

nal training set of 7,200 noisy sentences (3,600 per

language), exposing the SLM to a wide spectrum of

potential input corruptions.

Final Dataset Structure

The complete training dataset comprises 7,800 sam-

ples (3,900 per language), consisting of the 600 clean

base sentences and their 7,200 perturbed variants.

Each sample is an input-output pair for ﬁne-tuning

the SLM: the input is a (clean or perturbed) sentence

(e.g., "Inspeckt the dor panel..."), and the output is the

corresponding canonical defect term (e.g., deforma-

tion). This structure enables the SLM to learn robust

mappings from a wide range of corrupted inputs back

to the correct labels, equipping it to handle real-world

ASR transcription errors.

https://www.anthropic.com/news/claude-3-5-sonnet

5.2 Test Set

The test set has been constructed by generating 20

new, distinct sentences for each term in the two vo-

cabularies using the same Claude-Sonnet 3.5 LLM

employed for the training set base sentences, ensur-

ing no overlap between training and test samples.

These sentences were then converted into speech us-

ing Bark, a text-to-speech model

. We also publicly

share the link containing the text ﬁles used for the

training set used for ﬁne-tuning the SLM corrector

and the synthetic voice ﬁles of the test set for the eval-

uation step we performed

6 EXPERIMENTAL EVALUATION

6.1 Fine-Tuning

We carried out the ﬁne-tuning of the SLM using the

structured prompt shown in Figure 1 that instructs

the model to extract the correct defect term from a

noisy input sentence, given the closed vocabulary.

During training, each input sentence (clean or per-

turbed) and its target term were embedded into this

prompt. During inference, the ASR’s output is placed

into the same template, and the SLM generates the

corresponding vocabulary term. Fine-tuning was car-

ried out on an L4 GPU using the Unsloth frame-

work

. We employed the Quantization-aware Low-

Rank Adaptation (QLoRA)(Dettmers et al., 2023)

PEFT technique(Mangrulkar et al., 2022). The model

was trained for one epoch with a learning rate of 2e-4,

a weight decay of 0.001, and the AdamW 8-bit opti-

mizer.

6.2 Noise Injection for Realistic

Evaluation

To evaluate the pipeline’s robustness, we mixed the

clean test audio with real-world environmental noise

(e.g., car interior recordings). The Signal-to-Noise

Ratio (SNR) was controlled at two levels, represent-

ing moderate (5 dB) and high (-5 dB) interference, al-

lowing for an accurate evaluation of performance un-

der varying acoustic conditions.

https://huggingface.co/suno/bark

https://github.com/Marcomurgia97/audios_NATTER

_evaluation

https://unsloth.ai/

Noise-Robust Speech Transcription with Quantized Language Model Correction for Industrial Settings

481

Table 1: Examples of three defects.

(a) Deformation (b) Waviness (c) Glue Residue

Prompt

Given the transcribed voice phrase:

‘he deformaion is mre pronunced

towads th pane edges’ extract and

return only the correct term from the

following vocabulary:

[positive bulge, negative bulge,

waviness, deformation, reworking

traces, crack, failure, scratches,

glue residues, orange peel effect]

Note: There may be transcription

errors in the original phrase.

Identify the vocabulary term that best

matches the intended meaning in the

phrase, correcting any transcription

errors.

Return only the correct

term, without any additional

explanation.

Example: given the phrase ‘The

bidwark shows a dfomaton here’,

the correct term to return would be

‘deformation’. It may also happen that

there are words that are correct from a

lexical point of view but inappropriate

for the sentence. For example, given

the phrase ‘I can see a primitive

pulse’, the correct term to return

would be ‘positive bulge’.

Answer:‘Deformation’.

Figure 1: An example of a prompt used for ﬁne-tuning with

the target answer Deformation.

6.3 Evaluation Metrics and System

Variants

We evaluate our pipeline as a single-label multi-class

classiﬁcation task, where each defect term (e.g., de-

formation, waviness, see Table 1) is a distinct class.

To assess the impact of the SLM corrector, we com-

pare the performance of the pipeline with and without

it for both English and Italian. System performance

is measured using standard like Precision, Recall, and

F1-Score.

7 DISCUSSIONS

In this section, we analyze the pipeline’s performance

by comparing the baseline ASR output against the

results obtained with our SLM corrector. To assess

the performance-efﬁciency trade-off for resource-

constrained environments, we evaluate two ASR

model sizes: the lightweight Whisper Tiny (39M

parameters) and the more powerful Whisper Large

(1.5B parameters). The analysis is conducted for both

Italian (Tables 2, 3, 4) and English (Tables 5, 6, 7)

across three noise conditions: no added noise, SNR 5

dBW, and SNR -5 dBW. Performance is measured us-

ing Precision, Recall, and F1-Score, based on a strict

exact match criterion (case and singular/plural forms

are ignored).

The analysis of Italian language performance

(Tables 2-4) reveals two key ﬁndings. First, the

SLM corrector is crucial for the effectiveness of the

lightweight Whisper Tiny model. In noise-free con-

ditions, the SLM boosts Tiny’s F1-score from a poor

53.49 to an excellent 95.54, primarily by raising its

low Recall (42.50) to 95.50. This corrected per-

formance surpasses that of the uncorrected Whisper

Large (F1 91.64), establishing the Tiny+SLM combi-

nation as a computationally efﬁcient alternative. Sec-

ond, as noise increases, the impact of SLM becomes

even more pronounced. Under moderate noise (SNR

5 dBW), it mitigates Tiny’s performance drop, raising

its F1 from 44.16 to 90.53. In the most challenging

scenario (SNR -5 dBW), where the Tiny model is un-

effective (F1 34.45), the SLM corrector restores sys-

tem usability, achieving a robust F1-score of 87.04.

Whisper Large results show a similar trend, whose

F1-score is maintained at near-perfect levels by the

SLM (99.50 at no noise, 97.99 at -5 dBW). In all con-

ditions, the SLM contribution allows a signiﬁcant re-

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

482

Table 2: Italian Recognition Performance - No Noise.

Metric Global Defect Type

Bollo Negativo Bollo Positivo Buccia Di Arancia Buco Deformazione Grafﬁo Ondulazione Residuo Di Colla Rottura Traccia

Size: Tiny

F1 SLM 95.54 95.00 95.00 97.44 90.91 97.56 97.56 95.00 94.74 94.74 97.44

F1 Whisper 53.49 0.00 26.09 62.07 85.71 94.74 40.00 57.14 18.18 62.07 88.89

Precision SLM 95.88 95.00 95.00 100.00 83.33 95.24 95.24 95.00 100.00 100.00 100.00

Precision Whisper 90.00 0.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 95.50 95.00 95.00 95.00 100.00 100.00 100.00 95.00 90.00 90.00 95.00

Recall Whisper 42.50 0.00 15.00 45.00 75.00 90.00 25.00 40.00 10.00 45.00 80.00

Size: Large

F1 SLM 99.50 100.00 100.00 100.00 97.56 100.00 100.00 100.00 97.44 100.00 100.00

F1 Whisper 91.64 85.71 85.71 91.89 97.44 97.44 85.71 85.71 97.44 91.89 97.44

Precision SLM 99.52 100.00 100.00 100.00 95.24 100.00 100.00 100.00 100.00 100.00 100.00

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 99.50 100.00 100.00 100.00 100.00 100.00 100.00 100.00 95.00 100.00 100.00

Recall Whisper 85.00 75.00 75.00 85.00 95.00 95.00 75.00 75.00 95.00 85.00 95.00

Table 3: Italian Recognition Performance - SNR 5.

Metric Global Defect Type

Bollo Negativo Bollo Positivo Buccia Di Arancia Buco Deformazione Grafﬁo Ondulazione Residuo Di Colla Rottura Traccia

Size: Tiny

F1 SLM 90.53 92.31 90.91 92.68 79.17 97.56 91.89 82.35 85.71 97.44 95.24

F1 Whisper 44.16 0.00 0.00 62.07 78.79 88.89 40.00 46.15 0.00 40.00 85.71

Precision SLM 92.26 94.74 83.33 90.48 67.86 95.24 100.00 100.00 100.00 100.00 90.91

Precision Whisper 70.00 0.00 0.00 100.00 100.00 100.00 100.00 100.00 0.00 100.00 100.00

Recall SLM 90.50 90.00 100.00 95.00 95.00 100.00 85.00 70.00 75.00 95.00 100.00

Recall Whisper 34.50 0.00 0.00 45.00 65.00 80.00 25.00 30.00 0.00 25.00 75.00

Size: Large

F1 SLM 98.24 97.56 100.00 100.00 92.68 100.00 100.00 100.00 97.44 94.74 100.00

F1 Whisper 89.88 85.71 82.35 88.89 94.74 97.44 85.71 85.71 97.44 88.89 91.89

Precision SLM 98.57 95.24 100.00 100.00 90.48 100.00 100.00 100.00 100.00 100.00 100.00

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 98.00 100.00 100.00 100.00 95.00 100.00 100.00 100.00 95.00 90.00 100.00

Recall Whisper 82.00 75.00 70.00 80.00 90.00 95.00 75.00 75.00 95.00 80.00 85.00

Table 4: Italian Recognition Performance - SNR -5.

Metric Global Defect Type

Bollo Negativo Bollo Positivo Buccia Di Arancia Buco Deformazione Grafﬁo Ondulazione Residuo Di Colla Rottura Traccia

Size: Tiny

F1 SLM 87.04 92.68 87.18 82.05 70.83 100.00 87.80 88.89 85.71 92.31 82.93

F1 Whisper 34.45 0.00 0.00 26.09 51.85 88.89 40.00 33.33 0.00 33.33 70.97

Precision SLM 88.63 90.48 89.47 84.21 60.71 100.00 85.71 100.00 100.00 94.74 80.95

Precision Whisper 70.00 0.00 0.00 100.00 100.00 100.00 100.00 100.00 0.00 100.00 100.00

Recall SLM 86.50 95.00 85.00 80.00 85.00 100.00 90.00 80.00 75.00 90.00 85.00

Recall Whisper 25.00 0.00 0.00 15.00 35.00 80.00 25.00 20.00 0.00 20.00 55.00

Size: Large

F1 SLM 97.99 100.00 97.56 100.00 95.00 100.00 97.56 100.00 100.00 94.74 95.00

F1 Whisper 77.88 62.07 57.14 62.07 66.67 97.44 85.71 75.00 91.89 88.89 91.89

Precision SLM 98.05 100.00 95.24 100.00 95.00 100.00 95.24 100.00 100.00 100.00 95.00

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 98.00 100.00 100.00 100.00 95.00 100.00 100.00 100.00 100.00 90.00 95.00

Recall Whisper 66.00 45.00 40.00 45.00 50.00 95.00 75.00 60.00 85.00 80.00 85.00

covery of the Recall, preserving system functionality

even in severe noise.

The results for English (Tables 5-7) mirror the

trends observed for Italian, conﬁrming the SLM cor-

rector’s critical role. The baseline Whisper Tiny

starts with a higher F1-score than its Italian counter-

part (72.06 in no-noise conditions), but is still ham-

pered by low Recall (61.50). The SLM corrector el-

Noise-Robust Speech Transcription with Quantized Language Model Correction for Industrial Settings

483

Table 5: English Recognition Performance - No Noise.

Metric Global Defect Type

Negative Bulge Positive Bulge Orange Peel Effect Crack Deformation Scratch Waviness Glue Residue Failure Reworking Trace

Size: Tiny

F1 SLM 96.01 95.24 90.48 92.68 97.44 100.00 100.00 94.74 94.74 94.74 100.00

F1 Whisper 72.06 57.14 40.00 62.07 91.89 100.00 100.00 40.00 51.85 91.89 85.71

Precision SLM 96.78 90.91 86.36 90.48 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 95.50 100.00 95.00 95.00 95.00 100.00 100.00 90.00 90.00 90.00 100.00

Recall Whisper 61.50 40.00 25.00 45.00 85.00 100.00 100.00 25.00 35.00 85.00 75.00

Size: Large

F1 SLM 99.00 100.00 100.00 100.00 95.00 100.00 100.00 100.00 97.44 100.00 97.56

F1 Whisper 88.54 62.07 78.79 94.74 94.74 97.44 100.00 91.89 70.97 100.00 94.74

Precision SLM 99.02 100.00 100.00 100.00 95.00 100.00 100.00 100.00 100.00 100.00 95.24

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 99.00 100.00 100.00 100.00 95.00 100.00 100.00 100.00 95.00 100.00 100.00

Recall Whisper 81.50 45.00 65.00 90.00 90.00 95.00 100.00 85.00 55.00 100.00 90.00

Table 6: English Recognition Performance - SNR 5.

Metric Global Defect Type

Negative Bulge Positive Bulge Orange Peel Effect Crack Deformation Scratch Waviness Glue Residue Failure Reworking Trace

Size: Tiny

F1 SLM 94.22 95.24 92.68 95.24 88.37 97.44 95.00 97.44 88.89 91.89 100.00

F1 Whisper 68.56 46.15 40.00 57.14 82.35 100.00 97.44 40.00 51.85 91.89 78.79

Precision SLM 94.99 90.91 90.48 90.91 82.61 100.00 95.00 100.00 100.00 100.00 100.00

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 94.00 100.00 95.00 100.00 95.00 95.00 95.00 95.00 80.00 85.00 100.00

Recall Whisper 57.00 30.00 25.00 40.00 70.00 100.00 95.00 25.00 35.00 85.00 65.00

Size: Large

F1 SLM 98.00 100.00 97.56 97.44 92.68 100.00 100.00 100.00 94.74 100.00 97.56

F1 Whisper 84.91 51.85 66.67 88.89 94.74 97.44 94.74 91.89 70.97 100.00 91.89

Precision SLM 98.10 100.00 95.24 100.00 90.48 100.00 100.00 100.00 100.00 100.00 95.24

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 98.00 100.00 100.00 95.00 95.00 100.00 100.00 100.00 90.00 100.00 100.00

Recall Whisper 76.50 35.00 50.00 80.00 90.00 95.00 90.00 85.00 55.00 100.00 85.00

Table 7: English Recognition Performance - SNR -5.

Metric Global Defect Type

Negative Bulge Positive Bulge Orange Peel Effect Crack Deformation Scratch Waviness Glue Residue Failure Reworking Trace

Size: Tiny

F1 SLM 91.25 97.56 88.89 95.24 85.71 100.00 94.74 97.44 85.71 78.79 88.37

F1 Whisper 52.83 40.00 18.18 51.85 62.07 97.44 82.35 18.18 33.33 78.79 46.15

Precision SLM 93.06 95.24 80.00 90.91 81.82 100.00 100.00 100.00 100.00 100.00 82.61

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 91.00 100.00 100.00 100.00 90.00 100.00 90.00 95.00 75.00 65.00 95.00

Recall Whisper 40.50 25.00 10.00 35.00 45.00 95.00 70.00 10.00 20.00 65.00 30.00

Size: Large

F1 SLM 95.98 95.00 93.02 100.00 92.68 100.00 97.44 97.44 91.89 92.31 100.00

F1 Whisper 73.42 40.00 40.00 66.67 88.89 97.44 91.89 82.35 46.15 91.89 88.89

Precision SLM 96.72 95.00 86.96 100.00 90.48 100.00 100.00 100.00 100.00 94.74 100.00

Precision Whisper 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Recall SLM 95.50 95.00 100.00 100.00 95.00 100.00 95.00 95.00 85.00 90.00 100.00

Recall Whisper 62.50 25.00 25.00 50.00 80.00 95.00 85.00 70.00 30.00 85.00 80.00

evates its performance to a competitive 96.01, once

again making the Tiny+SLM setup a powerful, low-

resource alternative to the uncorrected Whisper Large

(F1 88.54). As noise levels increase, the SLM’s cor-

rective power ensures system robustness. Under mod-

erate noise (SNR 5 dBW), it lifts Tiny’s F1-score from

68.56 to 94.22. In the high-noise scenario (SNR -

5 dBW), where the uncorrected performance of Tiny

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

484

degrades signiﬁcantly (F1 52.83), the SLM provides

a substantial boost to 91.25, effectively mitigating the

noise-induced errors. For Whisper Large, the SLM

consistently maintains near-perfect accuracy, raising

its F1 from 73.42 to 95.98 in the -5 dBW condition.

As with Italian, the primary mechanism for this im-

provement is a signiﬁcant recovery in Recall (e.g., for

Tiny at -5 dBW, from 40.50 to 91.00), conﬁrming that

the Tiny+SLM combination is a reliable and efﬁcient

solution even under adverse conditions.

Our analysis consistently reveals three key trends

across both languages and all noise levels. First, as

expected, ASR performance degrades with increasing

noise. Second, the SLM corrector signiﬁcantly en-

hances performance in all conditions, especially noisy

ones. Third, and most importantly, the SLM nar-

rows the performance gap between Whisper Tiny and

Whisper Large, establishing the Tiny+SLM combina-

tion as a computationally efﬁcient alternative. For

instance, at SNR -5 dBW, the corrected Tiny model

processed samples in just 0.24 seconds, compared to

0.79 seconds for the corrected Large variant on our

test system (Intel i7, RTX 3070). This signiﬁcant re-

duction in computational implies also lower energy

consumption, a critical factor for battery-powered de-

vices on the factory ﬂoor. While our work establishes

a strong proof-of-concept, future work will address

further real-world deployment challenges, including

on-device integration and user interface design.

Our choice to use an SLM for post-correction,

rather than ﬁne-tuning the ASR model itself, is a prag-

matic one. Fine-tuning an ASR model requires exten-

sive and costly domain-speciﬁc audio data collection.

Our approach, which leverages a pre-trained ASR

and focuses adaptation on an SLM trained with eas-

ily generated synthetic text, is far more data-efﬁcient.

The strong performance gains, especially with Whis-

per Tiny, validate this strategy. This modularity also

allows for rapid adaptation to new domains by simply

retraining the SLM corrector with a new text dataset,

a much simpler task than acquiring new audio record-

ings. An interesting phenomenon was observed: for

some terms, the F1-score was higher under severe

noise (SNR -5 dBW) than moderate noise (SNR 5

dBW). This counterintuitive result likely stems from

the nature of the ASR errors. At -5 dBW, the highly

corrupted output may paradoxically produce error

patterns that more closely match the synthetic pertur-

bations in the SLM’s training data (see Section 5), en-

abling more effective correction. Conversely, errors

at 5 dBW, though milder, might be of a type less rep-

resented in our training set, thus limiting the SLM’s

corrective ability. This highlights the critical role of

the training data’s composition in the SLM’s effec-

tiveness at mitigating speciﬁc types of noise-induced

errors.

8 CONCLUSIONS

We presented a three-stage pipeline for robust speech

transcription in noisy industrial settings, combining

noise ﬁltering (DeepFilterNet), ASR (Whisper), and

a post-correction module. The core of our contribu-

tion is a lightweight ﬁne-tuned SLM (Gemma 2 2B-it)

that reﬁnes ASR outputs based on a domain-speciﬁc

defect ontology. Our results show that this SLM

corrector substantially improves F1-scores across all

tested conditions, mitigating noise-induced transcrip-

tion errors. It enables the lightweight Whisper Tiny to

achieve performance comparable to the much larger

Whisper Large variant, conﬁrming the pipeline’s suit-

ability for ofﬂine deployment on resource-limited

hardware. The evaluation was conducted on a syn-

thetic dataset, which, while allowing for rigorous test-

ing, does not capture the full complexity of real-world

scenarios. Key limitations include a lack of speaker

variability (e.g., accents, speaking rates) and environ-

mental noise diversity (e.g., impulsive sounds, over-

lapping speech). The strong performance observed

therefore establishes a promising proof-of-concept.

Our planned mitigation strategy is to conduct further

real-data validation in collaboration with our indus-

trial partner. This important next step, involving the

collection and testing of on-ﬁeld data, will be essen-

tial for bridging the gap between our results and a

practical deployment.

REFERENCES

Andreyev, A. (2025). Quantization for openai’s whisper

models: A comparative analysis.

Bain, M., Huh, J., Han, T., and Zisserman, A. (2023). Whis-

perx: Time-accurate speech transcription of long-

form audio. INTERSPEECH 2023, pages 4489–4493.

Bandela, S. R., Sharma Sadhu, S., Rathore, V. S., and

Jagini, S. K. (2023). Development of noise robust

automatic speech recognition system. In 2023 14th

International Conference on Computing Communica-

tion and Networking Technologies (ICCCNT), pages

1–6.

Bodepudi, A., Reddy, M., Gutlapalli, S., and Mandapuram,

M. (2019). Voice recognition systems in the cloud

networks: Has it reached its full potential? Asian

Journal of Applied Science and Engineering, 8:51–60.

Dettmers, T., Pagnoni, A., Holtzman, A., and Zettlemoyer,

L. (2023). Qlora: Efﬁcient ﬁnetuning of quantized

llms. In Oh, A., Neumann, T., Globerson, A., Saenko,

K., Hardt, M., and Levine, S., editors, Advances in

Noise-Robust Speech Transcription with Quantized Language Model Correction for Industrial Settings

485

Neural Information Processing Systems, volume 36,

pages 10088–10115. Curran Associates, Inc.

Dua, M., Akanksha, and Dua, S. (2023). Noise robust auto-

matic speech recognition: review and analysis. Int. J.

Speech Technol., 26(2):475–519.

Ephraim, Y. and Malah, D. (1984). Speech enhancement us-

ing a minimum-mean square error short-time spectral

amplitude estimator. IEEE Transactions on Acoustics,

Speech, and Signal Processing, 32(6):1109–1121.

G., T. Y., B.G., N., and Jayanna, H. (2024). Development

of noise robust real time automatic speech recognition

system for kannada language/dialects. Engineering

Applications of Artiﬁcial Intelligence, 135:108693.

Gholami, A., Kim, S., Zhen, D., Yao, Z., Mahoney, M., and

Keutzer, K. (2022). A Survey of Quantization Methods

for Efﬁcient Neural Network Inference, pages 291–

326.

Gomez, R. and Kawahara, T. (2010). Optimizing spec-

tral subtraction and wiener ﬁltering for robust speech

recognition in reverberant and noisy conditions. In

2010 IEEE International Conference on Acoustics,

Speech and Signal Processing, pages 4566–4569.

Huang, X., Baker, J., and Reddy, R. (2014). A historical

perspective of speech recognition. Commun. ACM,

57(1):94–103.

Jiang, Y. and Poellabauer, C. (2021). A sequence-to-

sequence based error correction model for medical au-

tomatic speech recognition. In 2021 IEEE Interna-

tional Conference on Bioinformatics and Biomedicine

(BIBM), pages 3029–3035.

Kamahori, K., Kasai, J., Kojima, N., and Kasikci, B. (2025).

Liteasr: Efﬁcient automatic speech recognition with

low-rank approximation.

Li, J., Deng, L., Gong, Y., and Haeb-Umbach, R. (2014).

An overview of noise-robust automatic speech recog-

nition. IEEE/ACM Transactions on Audio, Speech,

and Language Processing, 22(4):745–777.

Ma, R., Qian, M., Manakul, P., Gales, M., and Knill, K.

(2023). Can generative large language models per-

form asr error correction?

Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul,

S., and Bossan, B. (2022). Peft: State-of-the-art

parameter-efﬁcient ﬁne-tuning methods. https://gith

ub.com/huggingface/peft.

Mehrish, A., Majumder, N., Bharadwaj, R., Mihalcea, R.,

and Poria, S. (2023). A review of deep learning tech-

niques for speech processing. Information Fusion,

99:101869.

Moore, A., Peso Parada, P., and Naylor, P. (2017). Speech

enhancement for robust automatic speech recognition:

Evaluation using a baseline system and instrumental

measures. Computer Speech & Language, 46:574–

584.

Orel, D. and Varol, H. A. (2023). Noise-robust automatic

speech recognition for industrial and urban environ-

ments. In IECON 2023- 49th Annual Conference of

the IEEE Industrial Electronics Society, pages 1–6.

Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey,

C., and Sutskever, I. (2022). Robust speech recogni-

tion via large-scale weak supervision.

Rao, M., Raju, A., Dheram, P., Bui, B., and Rastrow, A.

(2020). Speech to semantics: Improve asr and nlu

jointly via all-neural interfaces. In Interspeech 2020,

page 876–880. ISCA.

Schröter, H., Escalante-B., A. N., Rosenkranz, T., and

Maier, A. (2022). DeepFilterNet: A low complexity

speech enhancement framework for full-band audio

based on deep ﬁltering. In ICASSP 2022 IEEE Inter-

national Conference on Acoustics, Speech and Signal

Processing (ICASSP). IEEE.

Suh, J., Na, I., and Jung, W. (2024). Improving domain-

speciﬁc asr with llm-generated contextual descrip-

tions. In Proc. Interspeech 2024, pages 1255–1259.

Virtanen, T., Plumbley, M. D., and Ellis, D. (2017). Com-

putational Analysis of Sound Scenes and Events.

Springer Publishing Company, Incorporated, 1st edi-

tion.

Yang, C.-H. H., Gu, Y., Liu, Y. C., Ghosh, S., Bulyko, I.,

and Stolcke, A. (2023). Generative speech recognition

error correction with large language models and task-

activating prompting.

Zhang, Z., Geiger, J., Pohjalainen, J., Mousa, A. E.-D., Jin,

W., and Schuller, B. (2018). Deep learning for envi-

ronmentally robust speech recognition: An overview

of recent developments. ACM Trans. Intell. Syst. Tech-

nol., 9(5).

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

486