Abstract: To address the problem that existing speech recognition methods have insufficient multi-scale global dependent features and local detail feature extraction in Amdo Tibetan speech recognition, leading to the inability to effectively distinguish homophones and near-tone characters, this paper proposes a method based on aggregated Conformer multi-scale features for Amdo Tibetan speech recognition. Firstly, speed perturbation and spectral enhancement are employed to enhance data diversity. Secondly, a multi-layer Conformer encoder is used to extract multi-scale speech features, aggregating global dependency and local detailed representations. An attention-based statistical pooling layer is adopted to obtain weighted means and standard deviations of these features, allowing differential weighting of global and local characteristics. Finally, a joint decoding approach using CTC, and an Attention mechanism is applied to perform speech recognition. Experimental results on the XBMU-AMDO31 Amdo Tibetan speech dataset show that the proposed model outperforms the baseline Conformer model by 36.68% in terms of performance, effectively recognizing homophones and near-tone characters in Amdo Tibetan speech.

Key words: Conformer, Amdo Tibetan, speech recognition, multi-scale features, feature aggregation