Three JD Tech AI Papers Shine at ICASSP 2021

Neural Kalman Filtering for Speech Enhancement

The paper proposes a neural Kalman filtering framework that integrates neural networks with optimal filter theory, training the Kalman filter weights via supervised learning. It builds a recurrent neural network‑based speech temporal model, predicts long‑term envelope and Wiener filter spectra, and combines them using analytically derived optimal weights. Experiments on Librispeech, PNL‑100Nonspeech‑Sounds, and MUSAN datasets show improvements in SNR gain, PESQ, and STOI over traditional UNET and CRNN‑based methods.

Improving Prosody Modeling with Cross‑Utterance BERT Embeddings for End‑to‑End Speech Synthesis

This work extracts cross‑sentence features using a pre‑trained BERT model and feeds them into an end‑to‑end speech synthesis system to enhance prosody. Two usage strategies are explored: concatenating all context embeddings as a single input, and applying attention between each phoneme and the sequence of context embeddings to obtain weighted prosody representations. Evaluations on Chinese and English audiobook datasets demonstrate more natural and expressive synthesized speech, with listeners preferring the proposed method over baseline models.

Conversational Query Rewriting with Self‑Supervised Learning

The paper introduces a self‑supervised approach for query rewriting in multi‑turn dialogue systems. By randomly deleting or replacing co‑occurring words in the user query, the model learns to reconstruct the original query using historical context, reducing reliance on annotated data. An enhanced model, Teresa, adds a keyword detection module based on TextRank and an intent‑consistency module that aligns intent distributions between original and rewritten queries, improving rewrite quality and preserving user intent.

Overall, JD Tech has contributed over 350 papers to top AI conferences such as AAAI, IJCAI, CVPR, KDD, NeurIPS, ICML, ACL, and ICASSP, winning 19 first‑place awards, and continues to drive advancements in speech, vision, and machine learning to support digital transformation across industries.