T. J. Tsai

Mengyi Shan, TJ Tsai

This paper tackles the problem of verifying the authenticity of speech recordings from world leaders. Whereas previous work on detecting deep fake or tampered audio focus on scrutinizing an audio recording in isolation, we instead reframe the problem and focus on cross-verifying a questionable recording against trusted references. We present a method for cross-verifying a speech recording against a reference that consists of two steps: aligning the two recordings and then classifying each query frame as matching or non-matching. We propose a subsequence alignment method based on the Needleman-Wunsch algorithm and show that it significantly outperforms dynamic time warping in handling common tampering operations. We also explore several binary classification models based on LSTM and Transformer architectures to verify content at the frame level. Through extensive experiments on tampered speech recordings of Donald Trump, we show that our system can reliably detect audio tampering operations of different types and durations. Our best model achieves 99.7% accuracy for the alignment task at an error tolerance of 50 ms and a 0.43% equal error rate in classifying audio frames as matching or non-matching.

TJ Tsai, Kevin Ji

This paper studies composer style classification of piano sheet music images. Previous approaches to the composer classification task have been limited by a scarcity of data. We address this issue in two ways: (1) we recast the problem to be based on raw sheet music images rather than a symbolic music format, and (2) we propose an approach that can be trained on unlabeled data. Our approach first converts the sheet music image into a sequence of musical "words" based on the bootleg feature representation, and then feeds the sequence into a text classifier. We show that it is possible to significantly improve classifier performance by first training a language model on a set of unlabeled data, initializing the classifier with the pretrained language model weights, and then finetuning the classifier on a small amount of labeled data. We train AWD-LSTM, GPT-2, and RoBERTa language models on all piano sheet music images in IMSLP. We find that transformer-based architectures outperform CNN and LSTM models, and pretraining boosts classification accuracy for the GPT-2 model from 46\% to 70\% on a 9-way classification task. The trained model can also be used as a feature extractor that projects piano sheet music into a feature space that characterizes compositional style.

Mengyi Shan, TJ Tsai

This paper studies the problem of automatically generating piano score following videos given an audio recording and raw sheet music images. Whereas previous works focus on synthetic sheet music where the data has been cleaned and preprocessed, we instead focus on developing a system that can cope with the messiness of raw, unprocessed sheet music PDFs from IMSLP. We investigate how well existing systems cope with real scanned sheet music, filler pages and unrelated pieces or movements, and discontinuities due to jumps and repeats. We find that a significant bottleneck in system performance is handling jumps and repeats correctly. In particular, we find that a previously proposed Jump DTW algorithm does not perform robustly when jump locations are unknown a priori. We propose a novel alignment algorithm called Hierarchical DTW that can handle jumps and repeats even when jump locations are not known. It first performs alignment at the feature level on each sheet music line, and then performs a second alignment at the segment level. By operating at the segment level, it is able to encode domain knowledge about how likely a particular jump is. Through carefully controlled experiments on unprocessed sheet music PDFs from IMSLP, we show that Hierarachical DTW significantly outperforms Jump DTW in handling various types of jumps.