As a fundamental technique in radionuclide identification and radiation monitoring, γ spectroscopy is a common method used in risk and hazard assessment studies. Analysis in γ spectroscopy often involves identification and classification of radioactive sources. The knowledge of known γ-emitting radioisotopes allows for the distinction to be made between different species by observing certain spectral features such as photopeaks due to the specific energies of the γ emissions, Compton continuum due to photon scattering, X-ray fluorescence, etc. Each radioisotope is uniquely identified based on these features, allowing a model to be developed for comparing new spectral data with existing data. Various machine learning techniques have been used and tested to observe the performance of different algorithms on γ-ray spectra, especially in regards to species classification. More recently, deep learning methods such as deep neural networks (DNNs) have been proven to be very successful in identifying and analysing γ-ray spectra, often by use of a combination of simulated and experimental data. These networks can classify radioisotope energy spectra with high precision. Being able to identify and also quantify contributions from isotopes in combination is challenging, especially where spectral features between multiple sources overlap, or the energy resolution is poor, causing further distortions in the spectrum. However, knowledge of the composition of a mixture of radiation sources is an ability which can be crucial in composition and elementary analysis from spectral information. This work sees the development of a regression-based convolutional neural network (CNN) which attempts to predict the sources and proportions of each source in simulated mixed-source spectra. A comparison between the network and a traditional library least-squares algorithm is also made. A Monte Carlo based simulation is used to produce the spectral data, using the GEANT4 software package, for 6 different isotopes and a basic experimental design modelled after a NaI(Tl) scintillation detector. A comprehensive dataset is generated for these isotopes, for use in future analysis works. With scintillators being a common choice for field work in radiation monitoring and similar environmental studies, a network with high performance and efficiency can provide a promising tool for automated spectral analysis and detection. The model is highly efficient, being able to process batches of spectra in a second, with accurate predictions having mean-square error on the order of ∼10