Jasper および QuartzNet ASR

The Jasper paper describes a deep temporal convolutional acoustic model with residual connections and a CTC training objective. The residual paths support a different route for information through the stack, without guaranteeing that gradients are perfect in every trained configuration. QuartzNet changes the convolutional design by separating temporal filtering from channel mixing. Its depthwise temporal part processes channels separately, and a pointwise part combines channel information. The number of parameters depends on the selected blocks and configuration. In the QuartzNet paper’s comparison table, JasperDR-10x5 is listed at about 333 million parameters and QuartzNet-15x5 at about 19 million; other QuartzNet variants have other counts. Reported LibriSpeech error rates likewise depend on the model, augmentation, and language-model decoding used, so those numbers should not be treated as universal accuracy parity. The authors tested augmentation choices and reported their best setup using speed perturbation combined with a cutout-style spectrogram method that removes small rectangles. This is an experiment-specific training choice, not a claim that every later model uses it. CTC addresses variable-length sequence labeling without manual frame-to-target alignment. Its path mapping merges adjacent repeated labels and removes blank symbols. For the stated path A, A, blank, A, the result is AA: the blank separates the two surviving A occurrences. Separable filters also support an explicit weight calculation. With kernel length three, four input channels, and eight output channels, depthwise plus pointwise weights total three times four plus four times eight, or 44, excluding biases and other layer parameters.