Source-linked AI summary
Boundary-Free Contextual Biasing: Depth-Adaptive Gating and Reading-Space Matching for Unsegmented Languages
Muhammad Huzaifah, Yu Pan, Zachary Yeo, Ningjie Bai, Guangzhao Yang
TL;DR
Japanese and Chinese lack the word boundaries and stable written units that existing contextual-biasing methods rely on. This paper uses a character automaton, depth-adaptive gating, and reading-space matching to bias frozen CTC models during greedy decoding. On Aishell-1 NE, two-stage biasing raises hard R1 recall to 66.5%, above trained CLAS at 64%.
Problem
Japanese and Chinese lack word boundaries and stable above-character units, complicating contextual matching when homophones can produce the wrong characters despite correct audio.
Source: §1 · p. 1Method
The decoder combines character-level match depth for adaptive gating with reading-space matching to bias frozen CTC models during greedy decoding.
Source: §1 · p. 1, §2.3 · p. 2Results
On Aishell-1 NE hard R1, two-stage biasing raises recall from 8.9% to 66.5%, above trained CLAS at 64%.
Source: §4.2 · p. 3Takeaways & Limitations
The paper releases the first open Japanese contextual-biasing evaluation set, constructed model-free from JSUT and Common Voice.
Source: §3 · p. 3Both biasing mechanisms assume dense scripts, and multi-character vocabularies still need the masked endpoint.
Source: §5 · p. 4
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Contextual biasing helps frozen ASR models recover rare, meaningful vocabulary during a single greedy CTC pass. Japanese and Chinese lack word boundaries and stable above-character units, so the paper introduces depth-adaptive gating and reading-space matching to steer decoding from available match evidence.
1. INTRODUCTION
- Japanese and Chinese provide neither CTC word-boundary tokens nor stable above-character units, while homophones can leave correct audio represented by the wrong characters.The paper identifies readings as the stable matching object, spanning token boundaries structurally.Source: §1 · p. 1
- A depth-adaptive gate controls bias strength from per-character match depth, while reading-space matching addresses homophone errors that surface matching cannot engage.The authors note that surface and reading matches tie in accuracy on their own; the gate uses depth to determine how hard to push.Source: §1 · p. 1
- Existing approaches rely on boundaries, training-time modules, model-internal access, or prompting; the authors identify no prior unmodified-checkpoint reading-space automaton.The paper also identifies the lack of an open Japanese contextual-biasing set.Source: §1 · p. 1
- The paper contributes an open Japanese benchmark and analyzes biasing across two architectures, two languages, and list-composition regimes.Its analysis examines script-dependent thresholds, list-scale weights, tokenizer-granularity caps, and settings robust across these regimes.Source: §1 · p. 1
2. METHOD
The method replaces word-boundary signals with character-level automaton state and match depth during greedy CTC decoding. A depth-adaptive gate selects between masked and two-stage biasing, while a reading-space automaton handles homophone matches and combines its evidence with surface matches.
2.1. Boundary-free surface biasing
- A character-level Aho-Corasick automaton tracks hotword matches, and precomputed token-level tables let greedy CTC decoding apply each token’s bias and advance its state.Failure links allow matches to begin at any character position; match-depth potential grows with the number of characters matched, and abandoning a partial match is free.Source: §2.1 · pp. 1–2, Figure 1 · p. 2
2.2. Depth-adaptive gating
- The depth threshold assigns masked biasing to entry matches and uncapped two-stage selection to deeper matches, balancing entry control against traversal protection.Thresholding uses Han-equivalent units on mixed-script Japanese lists; the paper reports a single threshold d⋆=3 there.Source: §2.2 · p. 2
2.3. Reading-space biasing
- A reading-space automaton advances on homophone tokens but spends bias only on the listed surface, then fuses reading and surface credits as max(s, r) + λ min(s, r).The reading paths are prepared offline, and the fusion discounts correlated evidence; λ ranges from pure max at 0 to pure sum at 1.Source: §2.3 · p. 2
3. OPEN JA BIASING BENCHMARK
The paper releases a model-free Japanese contextual-biasing evaluation set, addressing the absence of an open Japanese benchmark and concerns about evaluation-set-mined lists. Its construction uses annotated lexical spans, screened targets, and per-utterance distractor lists, with complementary JSUT and Common Voice data.
3. OPEN JA BIASING BENCHMARK
- The first open Japanese contextual-biasing evaluation set is built model-free from JSUT and Common Voice, unlike prior Japanese lists mined by decoding evaluation sets with the tested model.The authors say model-free construction avoids making the set easier for a particular system.Source: §3 · pp. 2–3
- Targets are annotated lexical spans, filtered and screened, then paired with distractors at list sizes N ∈ {100, 500, 1000, 2000} that do not occur in their utterance references.Targets are stratified by length and script; distractors are sampled to match each corpus’s target-length profile.Source: §3 · p. 3
- JSUT is small, read, and fully verified, while Common Voice contributes nearly two orders of magnitude more occurrences from crowd-sourced speech with a partially screened tail.The release includes per-utterance lists, span annotations, a verification suite, and a card; audio remains under the source corpora’s licenses.Source: §3 · p. 3, Table 1 · p. 3
4. EXPERIMENT AND RESULTS
Decode-only biasing on frozen CTC models improves contextual recall, with effectiveness depending on list composition, gate regime, and whether errors are orthographic. Reading-space matching and surface biasing address different misses, while larger lists expose precision costs from stronger pushes.
4.2. JA and ZH context-biasing results
- No configuration wins everywhere: two-stage loses precision and CER on WenetSpeech common entities already recovered at 88.6%, while top-K masking improves on unbiased CER (5.51%).On JA, the combined configuration leads or ties at 63.4% and 53.7% recall, with CV CER below unbiased decoding (18.06% vs 18.50%).Source: §4.2 · p. 3
- On OWSM, unchanged masked biasing raises recall from 40.3% to 53.4% on JSUT and from 41.1% to 62.4% on CV at P≥93.8%.Two-stage falls below unbiased decoding in both recall and CER.Source: §4.2 · p. 3
4.3. Regime analysis
- At d⋆=3, adaptive gating matches masked recall on katakana (69.5%) and reaches 44.0% on kanji, versus 43.6% and 42.8% for the endpoints.The threshold sweep is at matched weight; the two endpoints each degrade in the band the other protects.Source: §4.3 · p. 3, Figure 2 · p. 4
4.2. JA and ZH context-biasing results
- On Aishell-1 NE, two-stage biasing raises SenseVoice R1 recall from 8.9% to 66.5%, surpassing trained CLAS at 64% without parameter updates.Table 2 covers frozen SenseVoice-small, Aishell-1 NE, untuned WenetSpeech transfer, and the JA benchmark.Source: §4.2 · p. 3, Table 2 · p. 4
4.3. Regime analysis
- Over a twentyfold increase in list size, adaptive biasing at w=3 loses 10.5 recall points and reaches 80% precision, while w=2 configurations lose 3–6 points and stay above 95% precision.The passage attributes the faster decay in precision than recall to spurious partial matches and says push magnitude matters more than the gate.Source: §4.3 · p. 4, Figure 2 · p. 4
4.4. Reading-space matching and combination
- Among surface-decoding misses, readings remain present in 72% of Aishell NEdense occurrences, 31% on JSUT, and 18% on CV; reading-space matching recovers 31 of 32, all 22, and 131 of 353 occurrences, respectively.Its gains concentrate on short kanji targets and are absent on long katakana names where reading and surface channels coincide.Source: §4.4 · p. 4
- Surface and reading channels miss different occurrences, so combining them matches or improves on either alone across datasets; on JSUT, maximum fusion (λ=0) achieves 63.4% recall and 97.3% precision.On JSUT, this compares with 59.7% recall and 95.5% precision for the alternative.Source: §4.4 · p. 4
- On WenetSpeech, frequency damping cuts excess insertions from 714 to 304 and raises precision from 96.9% to 98.6%, at a cost of 0.2 recall points.On the rareentity Aishell list, it cuts excess insertions from 19 to 8.Source: §4.4 · p. 4
5. CONCLUSION
The paper presents boundary-free contextual biasing for frozen CTC models, combining depth-adaptive gating with reading-space matching. It reports gains on Chinese and Japanese evaluations and identifies dense scripts as a scope assumption requiring recalibration for other writing systems.
5. CONCLUSION
- Boundary-free contextual biasing uses character-match depth for gating and per-mora or syllable reading state to recover homophone errors in frozen CTC models.The approach combines a character automaton with surface and reading evidence.Source: §5 · p. 4
- For frozen CTC models, boundary-free contextual biasing reaches 66.5% R1 recall on Aishell-1 NE, above a trained baseline, transfers unretuned to WenetSpeech and OWSM, and gains 25 points on the released JA benchmark.Source: §5 · p. 4
- Both mechanisms assume dense scripts; Thai-like scripts may accrue depth credit too quickly, while phonemic orthographies leave less room for reading-space matching, so information-unit schedules need recalibration.Source: §5 · p. 4