In recent years, Voice Activity Detection (VAD) is significant for differentiating speech and non-speech in audio signals, specifically in noisy environments. In existing, traditional VAD methods consist of specific limitations such as noise interference and speaker variability. In this research, Improved Multi-Talker-Recurrent Neural Network Transducer (IMT-RNNT) is employed for detecting voices activities effectively. In this process, spectral subtraction technique is utilized for reducing background noise from audio signal for enhancing the clarity of speech signals. Moreover, minimum–maximum (min–max) normalization is employed for scaling audio signals to a uniform range. Further, Mel-Frequency Cepstral Coefficients (MFCC) with Discrete Wavelet Transform (DWT) is used for extracting acoustic features by decomposing MFCC into various frequency components which enables effective representation of both low- and high-frequency features. Then, the proposed IMT-RNNT is used for detecting voices effectively in dynamic environments by joining acoustic encoder and prediction network for handling overlapping speech from multiple talkers effectively. The proposed IMT-RNNT achieves better results in terms of F1-score (99.3%) and Detection Cost Function (DCF) of 0.3 compared to the existing Adaptive Attention Span Transformer (AAT).