PEEKING INTO THE FUTURE FOR CONTEXTUAL BIASING

Ramaneswaran Selvakumar, Cindy Tseng, Esung Kim, Vijendra Raj Apsingekar, Yun Tang

Samsung Research America, USA

Methodology

Figure 1. Overview of our proposed approach.

We propose a future-aware contextual biasing approach that enables the decoder to peek into the future by predicting multiple upcoming tokens in parallel. Instead of only predicting the next token, the model produces K future-token logits at each decoding step and directly leverages these predictions to score candidate entities from a dynamic bias list.

Each entity is aligned with the future-token predictions using its subword sequence, and a lightweight scoring network converts these aligned logits into an entity score. During inference, the model decodes from a unified search space that contains both static vocabulary tokens and dynamic bias-list entities, without requiring separate bias encoders or cross-attention modules.

Experimental Results

Model N = 0 (no-bias) N = 100 N = 200 N = 500
test-clean test-other test-clean test-other test-clean test-other test-clean test-other
Baseline (AED) 2.73 (17.52/2.27) 6.01 (32.34/5.07) - - - - - -
CLAS [10] 3.12 (15.56/2.73) 6.95 (28.26/6.19) 3.09 (15.56/2.70) 6.89 (27.81/6.14) 3.08 (15.11/2.71) 6.89 (27.59/6.15) 3.13 (15.75/2.74) 6.93 (27.92/6.18)
AED-MTP 2.58 (17.27/2.27) 6.00 (30.63/5.12) 2.58 (17.27/2.27) 6.00 (30.63/5.12) 2.58 (17.27/2.27) 6.00 (30.63/5.12) 2.58 (17.27/2.27) 6.00 (30.63/5.12)
Ours (λ = 1) 2.61 (17.78/2.13) 6.25 (32.40/5.32) 2.34 (10.98/2.07) 5.82 (21.85/5.24) 2.36 (11.24/2.08) 5.82 (21.85/5.24) 2.37 (11.49/2.09) 5.88 (22.68/5.27)
Ours (λ = 4.4) 2.61 (17.78/2.13) 6.25 (32.40/5.32) 2.27 (8.70/2.07) 5.64 (17.22/5.22) 2.28 (9.02/2.07) 5.63 (17.16/5.22) 2.30 (9.27/2.08) 5.64 (17.55/5.21)

Table 1. WER (B-WER/U-WER) results on Librispeech-960. Bold values represent the best result among the same bias list size N.

Citation 

@misc{selvakumar2025peekingfuturecontextualbiasing,
      title={Peeking Into The Future For Contextual Biasing}, 
      author={Ramaneswaran Selvakumar and Cindy Tseng and Eesung Kim and Vijendra Raj Apsingekar and Yun Tang},
      year={2025},
      eprint={2512.17657},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.17657}, 
}