<p>Telugu exhibits considerable regional variation across dialects such as Kostha Andhra, Rayalaseema, and Telangana, often undermining the effectiveness of dialect-independent speech technologies. This study presents a hybrid CNN–LSTM framework for the identification of Telugu dialects and introduces a dialect-balanced Telugu speech corpus compiled from YouTube sources and manual recordings. The corpus has undergone annotation and validation by native speakers from the three dialect regions, aiming to include a diverse array of speaker demographics and genuine recording environments. A standardized preprocessing pipeline, which includes noise reduction, silence removal, and amplitude normalization, is implemented to ensure data integrity and comparability. For modeling purposes, CNNs examine spatial-phonetic patterns in spectrogram representations, while LSTMs detect temporal-prosodic dependencies in MFCC sequences. This dual-feature approach mitigates the limitations present in each architecture. A comprehensive evaluation is conducted to compare our method against CNN, LSTM, and Multinomial Naïve Bayes baselines, employing metrics including Accuracy, Precision, Recall, F1-score, and ROC-AUC. The findings from three dialect classes demonstrate that the proposed hybrid model achieves an accuracy of 91.2% and an F1-score of 0.91, along with a stable ROC-AUC. This highlights its effectiveness in managing noise, class imbalance, and dialect overlap. Significant contributions comprise: (i) a validated, dialect-balanced Telugu speech corpus intended as a benchmark resource, (ii) a hybrid CNN–LSTM model that integrates spectral and sequential features for improved dialect discrimination, (iii) thorough benchmarking across multiple metrics against strong baseline models, and (iv) practical applicability as a framework for dialect-aware ASR and speech applications in low-resource Indian contexts, including voice assistants, educational tools, and call-center analytics.</p>