Automatic speech recognition (ASR) in industrial environments presents a considerable challenge due to high ambient noise, which significantly affects system performance. In addition, in most cases, these systems are evaluated under controlled conditions that are not representative of real-world situations. This study analyzes the impact of noise on different deep learning architectures for speech recognition, focusing on robustness in adverse acoustic conditions. The experiments used a corpus of 4,126 audio samples in Mexican Spanish, recorded by 86 speakers with variations in speech style and intensity. The dataset includes clean recordings, samples from real industrial environments, and samples mixed with industrial noise at signal-to-noise ratio (SNR) levels of 5 dB, 10 dB, 15 dB, and variable. Traditional deep neural networks (DNNs) and models integrating i-vectors were trained. A cross-validation phase was carried out, training each model with SNR conditions different from those of its original training to evaluate generalization to unknown acoustic environments. The results show that DNN architectures with i-vectors achieved a minimum Word Error Rate (WER) of 3.13%, outperforming traditional DNNs. The iV-A3 model offered the most consistent performance against noise, with 3.38%, 3.13%, and 3.43% in different scenarios. In fine-tuning, ASR4 (2.62%), ASR5 (3.61%), and ASR16 (3.83%) stood out, showing potential for adapting models to specific noise levels, albeit with some risk of overfitting. These findings highlight the importance of preprocessing and careful selection of acoustic models and DNN configurations to improve the robustness of ASR in noisy industrial environments.