Children with specific language impairment (SLI) experience persistent communication challenges that affect their development. Early detection is vital for effective intervention. This study applies large language model (LLM) architectures to automatically classify speech from children with and without SLI using the Speech Database of Typical and SLI Children (1,574 samples). Audio features were extracted via Hugging Face’s Auto Feature Extractor with Whisper and Wav2Vec2 as backbone models, then fine-tuned using Auto Model For Audio Classification. was conducted using accuracy, precision, recall, and F1-score, which showed that Whisper achieved perfect classification (1.0 across all metrics), outperforming Wav2Vec2 (accuracy = 0.9556). The findings highlight Whisper’s strong capability to capture fine-grained acoustic and linguistic patterns associated with SLI, demonstrating its potential for early screening and clinical deployment.