Scholarlink

<p>Cognitive computational systems demand architectures that integrate hierarchical representation, temporal dynamics and computational efficiency for human-like speech understanding in real-world conditions. Continuous speech recognition systems face a trade-off between deep acoustic modeling and real-time performance. To address this, we propose a framework combining WaveRNN and Capsule Networks (CapsNet). Our framework addresses the challenge of continuous speech recognition by combining efficient waveform processing with hierarchical pattern analysis. WaveRNN processes raw audio waveforms to classify broad phonemic categories (e.g., vowels, stops), learning temporal patterns invariant to speaker characteristics. Then, a hierarchical CapsNet refines these features via dynamic routing, explicitly modeling part-to-whole relationships among phonemes. An entropy-weighted fusion mechanism aggregates capsule outputs to resolve ambiguities in noisy or overlapping contexts. Evaluated on the TIMIT corpus (a standard benchmark for phoneme recognition), the system achieves a segment-level phoneme error rate (PER) of 4.0 %. Baseline systems such as Whisper, wav2vec&#xa0;2.0 and NVIDIA Canary were adapted to this phoneme-level setting by mapping their word-level outputs into phoneme sequences, ensuring fair comparison. The proposed architecture combines temporal modeling efficiency with hierarchical structure, enhancing robustness to phoneme-level ambiguities.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing continuous speech recognition with CapsNet and WaveRNN: a transfer learning approach

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Emna Bouhajeb,
  • Chiraz Jlassi,
  • Najet Arous

摘要

Cognitive computational systems demand architectures that integrate hierarchical representation, temporal dynamics and computational efficiency for human-like speech understanding in real-world conditions. Continuous speech recognition systems face a trade-off between deep acoustic modeling and real-time performance. To address this, we propose a framework combining WaveRNN and Capsule Networks (CapsNet). Our framework addresses the challenge of continuous speech recognition by combining efficient waveform processing with hierarchical pattern analysis. WaveRNN processes raw audio waveforms to classify broad phonemic categories (e.g., vowels, stops), learning temporal patterns invariant to speaker characteristics. Then, a hierarchical CapsNet refines these features via dynamic routing, explicitly modeling part-to-whole relationships among phonemes. An entropy-weighted fusion mechanism aggregates capsule outputs to resolve ambiguities in noisy or overlapping contexts. Evaluated on the TIMIT corpus (a standard benchmark for phoneme recognition), the system achieves a segment-level phoneme error rate (PER) of 4.0 %. Baseline systems such as Whisper, wav2vec 2.0 and NVIDIA Canary were adapted to this phoneme-level setting by mapping their word-level outputs into phoneme sequences, ensuring fair comparison. The proposed architecture combines temporal modeling efficiency with hierarchical structure, enhancing robustness to phoneme-level ambiguities.