Language identification is the front end of multilingual speech-processing tasks. The study aims to enhance the accuracy of language identification in complex acoustic environments by proposing a multi-scale feature extraction method. This method replaces the baseline feature extraction network with a multi-scale feature extraction network (SE-Res2Net-CBAM-BILSTM) to extract multi-scale features. A multilingual cocktail party dataset was simulated, and comparative experiments were conducted with various models. The experimental results show that the proposed model achieved language identification accuracies of 97.6% for an Oriental language dataset and 75% for a multilingual cocktail party dataset Furthermore, comparative experiments show that our model outperformed three other models in the accuracy, recall, and F1 values. Finally, a comparison of different loss functions shows that the model performance was better when using focal loss.

In multi-lingual, multi-speaker environments (e.g., international conference scenarios), speech, language, and background sounds can overlap. In real-world scenarios, source separation techniques are needed to separate target sounds. Downstream tasks, such as ASR, speaker recognition, speech recognition, VAD, etc., can be combined with speech separation tasks to gain a better understanding. Since most of the evaluation methods for monophonic separation are either single or subjective, this paper used the downstream recognition task as an overall evaluation criterion. Thus, the performance could be directly evaluated by the metrics of the downstream task. In this paper, we investigated a two-stage training scheme that combined speech separation and language identification tasks. To analyze and optimize the separation performance of single-channel overlapping speech, the separated speech was fed to a language identification engine to evaluate its accuracy. The speech separation model was a single-channel speech separation network trained with WSJ0-2mix. For the language identification system, we used an Oriental Language Dataset and a dataset synthesized by directly mixing different proportions of speech groups. The combined effect of these two models was evaluated for various overlapping speech scenarios. When the language identification network model was based on single-person single-speech frequency spectrum features, Chinese, Japanese, Korean, Indonesian, and Vietnamese had significantly improved recognition results over the mixed audio spectrum.