This paper presents a new fine-grained audio classification technique designed and experimented in the context of the LifeCLEF 2015 bird species identification challenge. Inspired by recent works on fine-grained image classification, we introduce a new match kernel based on the shared nearest neighbors of the low level audio features extracted at the frame level. To make such strategy scalable to the tens of millions of MFCC features extracted from the tens of thousands audio recordings of the training set, we used high-dimensional hashing techniques coupled with an efficient approximate nearest neighbors search algorithm with controlled quality. Further improvements are obtained by (i) using a sliding window for the temporal pooling of the raw matches (ii) weighting each low level feature according to the semantic coherence of its nearest neighbors. Results show the effectiveness of the proposed technique which ranked 2nd among the 7 research groups participating to the LifeCLEF bird challenge.