The Conformer is the standard backbone for End-to-End (E2E) Automatic Speech Recognition (ASR), yet its quadratic complexity hinders streaming applications. While chunk-wise schemes offer linear complexity, they often suffer from limited cross-chunk interaction. We propose Swichformer (Switch context powered chunk-wise conformer), which introduces a “chunk-wise context-switch” mechanism to facilitate broad information exchange. Swichformer employs three partitioning strategies: regular (local context), shifted (adjacent left context), and sequential sampling (global historical context). This fusion enables the model to capture rich, complementary contextual information while preserving the efficiency of linear computation. Extensive evaluations on AISHELL-1/2, HKUST, LibriSpeech, and TED-LIUM3 demonstrate that Swichformer outperforms the evaluated chunk-wise baseline models, with roughly 3%–8% relative reductions in CER or WER. Furthermore, it achieves competitive performance compared with representative models that have substantially larger parameter counts and use Transducer decoding, as well as quadratic-complexity models (e.g., a CTC-only CER of 5.94% versus a Transducer CER of 5.78% for Emformer, and a two-pass CER of 5.09% versus 5.25% for U2 on AISHELL-1), while maintaining linear computational efficiency. These results suggest that Swichformer is a promising and efficient chunk-wise Conformer variant that performs strongly among the evaluated chunk-wise baselines and is competitive with representative history-aware streaming baselines. Source code will be released.
Neural speech codecs that discretize continuous speech signals into token sequences have become a widely used paradigm for learning compact speech representations and supporting a broad range of downstream applications. However, existing approaches often find it difficult to substantially reduce the token count while preserving high-fidelity reconstruction, which in turn limits the efficiency of downstream models in practical deployments. To address this challenge, we propose PRISM, a progressive representation compression with content-adaptive information-density-guided segmentation and quantizer modulation that enables compression at a user-specified frame rate while adaptively selecting the number of active codebooks for quantization based on content, thereby reducing the output token rate. PRISM comprises two key stages: (i) it leverages latent variation score to quantify frame-level information density, and subsequently merges frames into segments accordingly; and (ii) it estimates the average information density of each segment and accordingly dynamically determines the number of codebooks to activate, yielding more efficient coding. We integrate PRISM into a neural speech codec and conduct extensive experiments. Comprehensive evaluations demonstrate that PRISM can achieve compression at externally specified frame rates while maintaining high reconstruction quality with only a small number of active codebooks, validating its effectiveness in delivering a superior quality–efficiency trade-off for variable-rate neural speech coding. Audio samples are available at: https://ethuil.github.io/PRISMdemo/.
This paper presents a novel acoustic feature representation method for the severity-level classification of dysarthria in speaker- and text-independent modes. The proposed method begins by finding the instantaneous energy distribution of the speech signal by computing short-term Fourier magnitude spectra (STFT-MS) with a small hop size (1 ms). By mimicking the human auditory system, the frequency axis of STFT-MS is divided into M subbands separated on the Mel scale. The average spectral magnitude in each subband is then calculated to represent the instantaneous energy profile of sound units. The energy dynamic in each subband is computed by processing the average subband magnitude through a digital accumulator, followed by computing the slope over an analysis frame. These energy dynamic patterns are further enhanced through logarithmic compression. The result is an M-dimensional feature vector for each analysis frame, referred to as the logarithmic compressed spectral averaged temporal slope (LC-SATS). The LC-SATS feature effectively captures spectro-temporal energy distribution patterns of dysarthric speech over short time frames. The long-term temporal dependencies of the LC-SATS feature are then captured by appending first- and second-order derivatives to it (LC-SATS–Δ–Δ2). The dysarthria severity classification systems are developed using the time delay neural network (TDNN) and the ECAPA-TDNN architecture (which incorporates emphasized channel attention, propagation, and aggregation) for enhanced representation of the temporal dependencies captured by the features. The performance of the LC-SATS and LC-SATS–Δ–Δ2 features for dysarthria severity level classification is evaluated in comparison with established feature sets, including Mel-frequency cepstral coefficients (MFCC), MFCC–Δ–Δ2, log filter bank energies (LFBE), openSMILE features, and self-supervised Hidden Unit BERT (HuBERT) embedding. Experiments conducted on the Universal Access Speech Corpus demonstrate that the LC-SATS–Δ–Δ2 feature outperforms all explored features. In the speaker-independent test scenario, LC-SATS–Δ–Δ2 achieved an accuracy of 78.56% and an F1 score of 70.63% using the ECAPA-TDNN.
