Boundary-Free Contextual Biasing: Depth-Adaptive Gating and Reading-Space Matching for Unsegmented Languages

Muhammad Huzaifah; Yu Pan; Zachary Yeo; Ningjie Bai; Guangzhao Yang

eess.ASOct 7, 2026

Boundary-Free Contextual Biasing: Depth-Adaptive Gating and Reading-Space Matching for Unsegmented Languages

Authors: Muhammad Huzaifah, Yu Pan, Zachary Yeo, Ningjie Bai, Guangzhao Yang

Organizations: R&D Team, Recho Inc., Tokyo, Japan

Abstract

Contextual biasing supplies an ASR system with a list of expected words at inference time, but existing methods rely on word boundaries that Japanese and Chinese do not provide. We present a boundary-free biasing decoder for frozen public CTC models, built on a character-level Aho-Corasick automaton, with no training and no second pass. Two evidence-based mechanisms replace the boundary: a depth-adaptive gate that sets how hard to push from match depth, and reading-space matching for when the audio is right but the characters are wrong. On Aishell-1 NE's hard R1 subset we reach 66.5% recall, above the trained CLAS baseline (64%), transferring to WenetSpeech and to a second architecture without retuning. We release the first open Japanese contextual-biasing benchmark, where biasing lifts rare-word recall by 25 points at precision above 97%, and still by 19 and 22 points against 1,000-word lists.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Character-level Aho-Corasick automaton for 東京 and 京都; solid arrows are trie edges, the dashed one a failure link. Circled numbers trace the decoded stream 東京都 along the thick path, the failure link moving from the end of 東京 into the middle of 京都 without consuming a character, so both are matched. (b) The precomputed token-level table and per-frame decode flow; cells give the state reached and the bias $\psi(d^{\prime})-\psi(d)$ at $w{=}1$ , plus a lump $c$ per hotword completed.

corpus	utts	w/ target	hotwords	occurrences
JSUT	5,000	422 (8.44%)	477	489
CV ja	300,077	25,876 (8.62%)	4,357	27,108

Table 1: JA biasing benchmark statistics. JSUT is fully human-verified; CV is partly human-reviewed, the unreviewed residual noise rate is measured by a seeded blind sample and reported in the dataset card.

method	config	CER	Rec. all/R1	P	CER	Rec.	P	config	CER	Rec.	P	CER	Rec.	P
		Aishell-1 NE			WenetSpeech				JSUT			Common Voice ja
greedy	–	7.54	52.4/8.9	99.6	5.62	88.6	99.7	–	8.42	38.0	100	18.50	29.1	99.9
+ masked	$w{=}2.5$	5.70	77.4/53.1	98.8	5.51	93.9	98.8	$w{=}2$	8.52	59.7	99.3	17.77	53.9	99.7
+ two-stage	$w{=}4$	5.91	83.5/66.5	96.9	6.43	94.5	93.5	$w{=}2$	8.47	54.6	99.6	17.62	50.5	99.8
+ adaptive	$w{=}3$ , $d^{\star}{=}2$	5.48	81.0/60.6	98.7	5.69	94.1	98.3	$w{=}2$ , $d^{\star}{=}3$	8.52	60.3	99.3	17.69	53.5	99.7
+ reading	$w{=}2$	5.20	79.0/58.1	98.9	5.41	93.5	98.6	$w{=}1.5$	8.68	57.5	98.0	18.41	42.2	99.5

Table 2: Decode-only biasing on frozen SenseVoice-small: Aishell-1 NE test, recall over the full list and over the hard R1 subset (trained references, R1 recall [ 15 ] : CLAS 64, SeACo 80, +ASF 87); its untuned transfer to WenetSpeech [ 19 ] ; the JA benchmark (§3; N=100, configs frozen on internal benchmark).

K{=}10

; P=biased-word precision. Reading and combined (adaptive+reading) rows are frequency-damped (§2.3).

Figure 2: Gate-family regimes. (a) Threshold sweep at matched weight (

w{=}2

N{=}1000

); each endpoint degrades in the band the other protects. (b) Effect of distractor list-size scaling on recall.

Explore similar work

BaLEEN: Biasing with Latent Encoded Entities for Context-Aware ASR
Sep 29, 2026Chihiro Taguchi, Yotaro Kubo, Rujikorn CharakornDomain Adaptation for ASR Automatic Speech Recognition
University of Notre Dame, Department of Computer Science and Engineering, IN, USA · Sakana AI, Tokyo, Japan
How to Recognize New Words: A Comparison Between Context Biasing Methods and Speech LLMs
Aug 6, 2026Christian Huber, Alexander WaibelASR Evaluation Automatic Speech Recognition
Interactive Systems Lab · Karlsruhe Institute of Technology · Karlsruhe, Germany +2
PTC-Bias: Phoneme-Level Temporal Competition for Bias Retrieval and Post-Decoding Correction in Speech LLMs
Sep 23, 2026Zhiqi Ai, Han Cheng, Shiyi Mu +2Automatic Speech Recognition Speech Language Models
Shanghai University, Shanghai, China · Xi’an Jiaotong-Liverpool University, Suzhou, China

Sep 29, 2026, cs.CLJ/K move · Enter open · S save

BaLEEN: Biasing with Latent Encoded Entities for Context-Aware ASR

Chihiro Taguchi, Yotaro Kubo, Rujikorn Charakorn

University of Notre Dame, Department of Computer Science and Engineering, IN, USA · Sakana AI, Tokyo, Japan

Domain Adaptation for ASR Automatic Speech Recognition

Open paper AskPDF arXiv

Transcribing domain-specific entities and rare proper nouns remains a major challenge in automatic speech recognition (ASR). In this paper, we propose BaLEEN (Biasing with Latent Encoded Entities), a lightweight, hypernetwork-based framework for dynamic contextual adaptation without fine-tuning the underlying ASR model. BaLEEN encodes variable-length contextual keywords using a pretrained language model, compresses them into a fixed sequence of latent vectors via a Perceiver bottleneck, and injects context-dependent bias vectors directly into the intermediate encoder representations of the ASR model. Because both the language model and the backbone ASR model remain entirely frozen during training, BaLEEN operates as a plug-and-play adapter that incurs zero computational overhead at inference time when context biases are precomputed. We evaluate our method on a CTC-based ASR model using a Wikipedia-derived corpus with annotated named entities and synthetic speech. Experimental results demonstrate that BaLEEN reduces keyword miss rate by 8.7% on the test set relative to the unbiased baseline while simultaneously improving overall word error rate by 21% and character error rate by 28%.

Abstract

Explore similar work

BaLEEN: Biasing with Latent Encoded Entities for Context-Aware ASR

How to Recognize New Words: A Comparison Between Context Biasing Methods and Speech LLMs

PTC-Bias: Phoneme-Level Temporal Competition for Bias Retrieval and Post-Decoding Correction in Speech LLMs