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

# PTC-Bias: Phoneme-Level Temporal Competition for Bias Retrieval and Post-Decoding Correction in Speech LLMs
Source: [https://arxiv.org/html/2609.28727](https://arxiv.org/html/2609.28727)
Han ChengShiyi MuYongjin ZhouShugong Xu††thanks:\*Corresponding authors

###### Abstract

Contextual biasing improves rare\-word recognition in speech large language models \(SpeechLLMs\), but efficiently exploiting large bias lists remains challenging\. We proposePTC\-Bias, a two\-stage framework based on phoneme\-level temporal competition\. At the prefill stage,PTC Retrievalperforms frame\-synchronous phoneme decoding and temporal competition among candidate pronunciations, producing a compact bias\-word shortlist and corresponding speech intervals\. After SpeechLLM decoding,PTC Correctionconducts a second local competition between the retrieved candidates and mismatched transcript spans within these intervals\. Selective correction reduces near\-homophone and word\-segmentation errors while preserving correct transcriptions\. Both stages share the same phoneme posteriors and require no additional SpeechLLM forward pass\. Experiments on LibriSpeech show consistent gains across two SpeechLLMs and bias lists of up to 2000 words\. With Prompt\-SLAM\-ASR\-7B and 2000 bias words, PTC\-Bias reduces B\-WER by 23\.4%/23\.9% relative to CTC\-Filter on test\-clean/test\-other, while keeping U\-WER nearly unchanged\.

###### Index Terms:

speech large language models, contextual biasing, phoneme CTC, temporal competition, post\-decoding correction

††address:1Shanghai University, Shanghai, China
2Xi’an Jiaotong\-Liverpool University, Suzhou, China## 1Introduction

Speech large language models \(SpeechLLMs\) have achieved strong performance in automatic speech recognition \(ASR\), but they still struggle with rare words such as names, locations, and technical terms\. Contextual ASR addresses this problem by incorporating a user\-provided bias list through deep biasing, text\-based adaptation, or contextual prompting\[[12](https://arxiv.org/html/2609.28727#bib.bib1),[10](https://arxiv.org/html/2609.28727#bib.bib2),[17](https://arxiv.org/html/2609.28727#bib.bib3),[7](https://arxiv.org/html/2609.28727#bib.bib4)\]\. For SpeechLLMs, however, directly inserting a large bias list into the prompt increases inference cost and may introduce lexical interference, although only a small fraction of the list is usually relevant to a given utterance\. This has motivated retrieval\-based approaches that select a compact set of acoustically relevant bias words before SpeechLLM decoding\[[20](https://arxiv.org/html/2609.28727#bib.bib5),[13](https://arxiv.org/html/2609.28727#bib.bib6),[8](https://arxiv.org/html/2609.28727#bib.bib7)\]\.

Existing bias\-word selection methods rely on phonetic similarity, speech–text representations, intermediate ASR hypotheses, or learned retrieval scores\[[20](https://arxiv.org/html/2609.28727#bib.bib5),[13](https://arxiv.org/html/2609.28727#bib.bib6),[8](https://arxiv.org/html/2609.28727#bib.bib7),[9](https://arxiv.org/html/2609.28727#bib.bib8),[11](https://arxiv.org/html/2609.28727#bib.bib9),[6](https://arxiv.org/html/2609.28727#bib.bib10)\]\. Embedding\-based methods scale efficiently to large bias lists but often lack precise temporal localization, whereas hypothesis\-based methods may miss a relevant word if it is absent from the initial transcription\. Moreover, most methods rank candidates independently, allowing multiple near\-homophones associated with the same speech segment to be selected together rather than compared directly\.

Phoneme\-based keyword spotting offers finer frame\-level evidence\. CTC\-based word spotters search candidate pronunciations over phoneme posteriors\[[4](https://arxiv.org/html/2609.28727#bib.bib11),[14](https://arxiv.org/html/2609.28727#bib.bib12)\], while recent user\-defined keyword\-spotting systems employ streaming phoneme search, multi\-stage matching, or frame\-asynchronous decoding\[[3](https://arxiv.org/html/2609.28727#bib.bib14),[19](https://arxiv.org/html/2609.28727#bib.bib15),[2](https://arxiv.org/html/2609.28727#bib.bib16)\]\. However, these methods are primarily designed to detect individual keywords and typically evaluate each candidate through a separate decoding or matching path, without explicitly modeling competition between overlapping pronunciations\.

Retrieval alone is also insufficient for reliable contextual correction\. Even when the correct bias word is included in the prompt, a SpeechLLM may still output a frequent near\-homophone or produce an incorrect word segmentation\. Existing word\-spotting approaches can insert detected keywords into ASR hypotheses\[[4](https://arxiv.org/html/2609.28727#bib.bib11),[1](https://arxiv.org/html/2609.28727#bib.bib13)\], but a false detection may lead to an unnecessary replacement of an already correct transcription\. Reliable correction therefore requires a second comparison between the retrieved bias word and the corresponding transcript span, using acoustic evidence localized to the detected speech interval\.

To address these limitations, we proposePTC\-Bias111Project code:[https://github\.com/aizhiqi\-work/PTC\-Bias](https://github.com/aizhiqi-work/PTC-Bias), a two\-stage contextual biasing framework based on phoneme\-level temporal competition\. During prefill,PTC Retrievalperforms frame\-synchronous phoneme decoding and temporal competition among candidate pronunciations, producing a compact bias\-word shortlist together with their speech intervals\. After SpeechLLM decoding,PTC Correctionreuses the detected intervals and cached phoneme posteriors to conduct a second local competition between each retrieved candidate and its mismatched transcript span\. The two stages thus use the same phoneme\-level acoustic evidence and require no additional SpeechLLM forward pass\. Experiments on LibriSpeech demonstrate consistent improvements over the corresponding SpeechLLM and CTC\-Filter baselines across two SpeechLLMs and bias\-list sizes ranging from 100 to 2000\. With Prompt\-SLAM\-ASR\-7B and 2000 bias words, PTC\-Bias reduces B\-WER by 23\.4% and 23\.9% relative to CTC\-Filter on test\-clean and test\-other, respectively, while keeping U\-WER nearly unchanged\.

## 2Proposed Method

![Refer to caption](https://arxiv.org/html/2609.28727v1/overview-hotword-cropped.png)Figure 1:Overview of PTC\-Bias\. A lightweight phoneme\-CTC branch extracts frame\-level phoneme posteriors from the frozen audio encoder\. PTC Retrieval performs temporal competition to select acoustically supported bias words and locate their speech intervals before SpeechLLM decoding\. PTC Correction then reuses the intervals and cached posteriors to correct mismatched transcript spans through local competition\.### 2\.1Overview

As shown in Fig\.[1](https://arxiv.org/html/2609.28727#S2.F1), PTC\-Bias uses a lightweight phoneme\-CTC branch to obtain frame\-level phoneme posteriors\. Before decoding, PTC Retrieval selects bias words and their speech intervals via temporal competition for prompting; after decoding, PTC Correction uses the same posteriors to correct mismatched transcript spans within those intervals\. Both stages share the phoneme posteriors and require no additional audio\-encoder or SpeechLLM forward pass\.

### 2\.2Contextual SpeechLLM

The Contextual SpeechLLM comprises an audio encoder, an audio projection module, and an autoregressive LLM decoder\. The projected speech representations and a textual bias prompt are jointly provided to the LLM for transcription\. Since a large bias list increases the prompt length and may introduce lexical interference\[[8](https://arxiv.org/html/2609.28727#bib.bib7)\], we include at mostKKwords selected by PTC Retrieval \(Sec\.[2\.4](https://arxiv.org/html/2609.28727#S2.SS4)\) and retain their speech intervals for subsequent verification\. This preserves the original SpeechLLM interface without modifying its parameters\. Remaining errors are handled by PTC Correction \(Sec\.[2\.5](https://arxiv.org/html/2609.28727#S2.SS5)\)\.

### 2\.3Phoneme Decoding

We attach a lightweight phoneme\-CTC branch to the frozen audio encoder\. Its input is a learned weighted sum of allLLTransformer\-layer outputs:

𝐇¯t=∑ℓ=1Lαℓ​𝐇t\(ℓ\),𝜶=softmax⁡\(𝐠\),\\overline\{\\mathbf\{H\}\}\_\{t\}=\\sum\_\{\\ell=1\}^\{L\}\\alpha\_\{\\ell\}\\mathbf\{H\}^\{\(\\ell\)\}\_\{t\},\\qquad\\boldsymbol\{\\alpha\}=\\operatorname\{softmax\}\(\\mathbf\{g\}\),\(1\)where𝐇t\(ℓ\)\\mathbf\{H\}^\{\(\\ell\)\}\_\{t\}is the representation at framettfrom layerℓ\\ell, andαℓ\\alpha\_\{\\ell\}is its normalized weight\. A lightweight Transformer and phoneme\-CTC head then produce frame\-level phoneme posteriors\. Reference transcripts and bias words are converted into stress\-marked ARPAbet sequences using g2pE\[[16](https://arxiv.org/html/2609.28727#bib.bib17)\], providing a shared phoneme vocabulary for training and retrieval\. Only the fusion weights, lightweight Transformer, and CTC head are optimized\. The resulting posteriors are computed once and shared by PTC Retrieval and PTC Correction\.

Table 1:Contextual ASR results \(%\) on LibriSpeech with different numbers of distractors\. Each entry reports test\-clean / test\-other\. In the Qwen3\-ASR block,†\\daggerdenotes WavLM\-based phoneme posteriors; otherwise, the AuT front\-end is used\. Lower is better\.Algorithm 1PTC RetrievalInput:Phoneme posteriors𝐏1:T\\mathbf\{P\}\_\{1:T\}; bias listℬ\\mathcal\{B\}

Output:Retrieved events

ℛ\\mathcal\{R\}
𝒲←G2P​\(ℬ\)\\mathcal\{W\}\\leftarrow\\textsc\{G2P\}\(\\mathcal\{B\}\);

𝒢←BuildTrie​\(𝒲\)\\mathcal\{G\}\\leftarrow\\textsc\{BuildTrie\}\(\\mathcal\{W\}\);

𝒟←InitializeCTC​\(𝒢\)\\mathcal\{D\}\\leftarrow\\textsc\{InitializeCTC\}\(\\mathcal\{G\}\);

ℰ←∅\\mathcal\{E\}\\leftarrow\\varnothing;

for*t←1t\\leftarrow 1toTT*do

𝒟←TrieCTCStep​\(𝒟,𝐏t\)\\mathcal\{D\}\\leftarrow\\textsc\{TrieCTCStep\}\(\\mathcal\{D\},\\mathbf\{P\}\_\{t\}\);

ℰ←UpdateBestEvents​\(ℰ,𝒟,t,β\)\\mathcal\{E\}\\leftarrow\\textsc\{UpdateBestEvents\}\(\\mathcal\{E\},\\mathcal\{D\},t,\\beta\);

ℰM←TopM​\(ℰ,M\)\\mathcal\{E\}\_\{M\}\\leftarrow\\textsc\{TopM\}\(\\mathcal\{E\},M\);

\{qk\}←IntervalFB​\(ℰM,ρ,τ\)\\\{q\_\{k\}\\\}\\leftarrow\\textsc\{IntervalFB\}\(\\mathcal\{E\}\_\{M\},\\rho,\\tau\);

ℛ←TopKByMarginal\(\{\(ξk,qk\):ξk∈ℰM,qk≥η\},K\)\\mathcal\{R\}\\leftarrow\\textsc\{TopKByMarginal\}\\bigl\(\\\{\(\\xi\_\{k\},q\_\{k\}\):\\xi\_\{k\}\\in\\mathcal\{E\}\_\{M\},\\ q\_\{k\}\\geq\\eta\\\},K\\bigr\);

return*ℛ\\mathcal\{R\}*;

### 2\.4PTC Retrieval

PTC Retrieval converts the bias words into stress\-marked phoneme sequences and compiles them into a shared prefix trie\. Following CTC\-based word spotting and streaming phoneme search\[[4](https://arxiv.org/html/2609.28727#bib.bib11),[14](https://arxiv.org/html/2609.28727#bib.bib12),[19](https://arxiv.org/html/2609.28727#bib.bib15),[2](https://arxiv.org/html/2609.28727#bib.bib16)\], it searches the phoneme posteriors frame by frame\. Each candidate may start at any frame, while candidates sharing a phoneme prefix reuse the same decoding states\. Algorithm[1](https://arxiv.org/html/2609.28727#algorithm1)summarizes the procedure\.

For a bias wordvkv\_\{k\}with pronunciation𝐰\(k\)\\mathbf\{w\}^\{\(k\)\}of lengthUkU\_\{k\}, the search retains its best local CTC path and corresponding interval\[ak,bk\]\[a\_\{k\},b\_\{k\}\]\. The length\-calibrated score is

Sk=max1≤a≤b≤T𝒞\(𝝅a:b\)=𝐰\(k\)\[1Uk∑t=ablogPt\(πt\)\+βlogUk\],S\_\{k\}=\\max\_\{\\begin\{subarray\}\{c\}1\\leq a\\leq b\\leq T\\\\ \\mathcal\{C\}\(\\boldsymbol\{\\pi\}\_\{a:b\}\)=\\mathbf\{w\}^\{\(k\)\}\\end\{subarray\}\}\\left\[\\frac\{1\}\{U\_\{k\}\}\\sum\_\{t=a\}^\{b\}\\log P\_\{t\}\(\\pi\_\{t\}\)\+\\beta\\log U\_\{k\}\\right\],\(2\)where𝒞\\mathcal\{C\}is the CTC collapse operator,Pt​\(πt\)P\_\{t\}\(\\pi\_\{t\}\)is the posterior probability of CTC symbolπt\\pi\_\{t\}at framett, andβ\\betacontrols length calibration\. The maximizing path determines the interval\[ak,bk\]\[a\_\{k\},b\_\{k\}\]\.

Each bias word contributes its best eventξk=\(vk,\[ak,bk\],Sk\)\\xi\_\{k\}=\(v\_\{k\},\[a\_\{k\},b\_\{k\}\],S\_\{k\}\)\. We retain theMMhighest\-scoring events asℰM\\mathcal\{E\}\_\{M\}and treat events with overlapping intervals as competing explanations of the same acoustic evidence\. LetℭM\\mathfrak\{C\}\_\{M\}denote all subsets ofℰM\\mathcal\{E\}\_\{M\}whose events have non\-overlapping intervals\. The empty subset is included to allow no bias word to be selected\. The weight of a compatible event set𝒜∈ℭM\\mathcal\{A\}\\in\\mathfrak\{C\}\_\{M\}is

ψ⁡\(𝒜\)=exp⁡\(∑ξj∈𝒜Sj−ρτ\),\\psi\(\\mathcal\{A\}\)=\\exp\\left\(\\sum\_\{\\xi\_\{j\}\\in\\mathcal\{A\}\}\\frac\{S\_\{j\}\-\\rho\}\{\\tau\}\\right\),\(3\)whereρ\\rhois the per\-event reference score andτ\\tauis the competition temperature\. The competition\-adjusted confidence of eventξk\\xi\_\{k\}is its inclusion marginal

qk=∑𝒜∈ℭMξk∈𝒜ψ⁡\(𝒜\)∑𝒜∈ℭMψ⁡\(𝒜\)\.q\_\{k\}=\\frac\{\\displaystyle\\sum\_\{\\begin\{subarray\}\{c\}\\mathcal\{A\}\\in\\mathfrak\{C\}\_\{M\}\\\\ \\xi\_\{k\}\\in\\mathcal\{A\}\\end\{subarray\}\}\\psi\(\\mathcal\{A\}\)\}\{\\displaystyle\\sum\_\{\\mathcal\{A\}\\in\\mathfrak\{C\}\_\{M\}\}\\psi\(\\mathcal\{A\}\)\}\.\(4\)These marginals are computed efficiently using interval forward–backward\. Events satisfyingqk≥ηq\_\{k\}\\geq\\etaare ranked byqkq\_\{k\}, and at mostKKbias words, together with their intervals and retrieval confidences, are passed to the Contextual SpeechLLM\.

### 2\.5PTC Correction

A retrieved bias word may still be transcribed as a near\-homophone or split into multiple words\. PTC Correction aligns the initial transcript𝐲^\\hat\{\\mathbf\{y\}\}to the cached phoneme posteriors and identifies a short, phonetically similar span𝐲^i:j\\hat\{\\mathbf\{y\}\}\_\{i:j\}overlapping each retrieved interval\[ak,bk\]\[a\_\{k\},b\_\{k\}\]\. To compare the retrieved wordvkv\_\{k\}with this span under the same phonetic context, we form𝐳k\+=𝐜L⊕G2P⁡\(vk\)⊕𝐜R\\mathbf\{z\}\_\{k\}^\{\+\}=\\mathbf\{c\}\_\{L\}\\oplus\\operatorname\{G2P\}\(v\_\{k\}\)\\oplus\\mathbf\{c\}\_\{R\}and𝐳k−=𝐜L⊕G2P\(𝐲^i:j\)⊕𝐜R\\mathbf\{z\}\_\{k\}^\{\-\}=\\mathbf\{c\}\_\{L\}\\oplus\\operatorname\{G2P\}\(\\hat\{\\mathbf\{y\}\}\_\{i:j\}\)\\oplus\\mathbf\{c\}\_\{R\}, where𝐜L\\mathbf\{c\}\_\{L\}and𝐜R\\mathbf\{c\}\_\{R\}are shared context phonemes\.

Both sequences are scored within a local windowΩk\\Omega\_\{k\}extended by five posterior frames on each side\. Their acoustic margin is

Δk=F⁡\(𝐳k\+,Ωk\)−F⁡\(𝐳k−,Ωk\),\\Delta\_\{k\}=F\(\\mathbf\{z\}\_\{k\}^\{\+\};\\Omega\_\{k\}\)\-F\(\\mathbf\{z\}\_\{k\}^\{\-\};\\Omega\_\{k\}\),\(5\)whereFFdenotes the CTC log forward probability\. A replacement requiresΔk\\Delta\_\{k\}to exceed a threshold and pass lexical and boundary checks\. Exact homophones, which phoneme\-CTC cannot distinguish, are handled only under restricted rare\-word or segmentation conditions\. Conflicting edits are resolved usingΔk\\Delta\_\{k\}and the retrieval marginalqkq\_\{k\}from Eq\. \([4](https://arxiv.org/html/2609.28727#S2.E4)\)\. All comparisons reuse cached posteriors without an additional audio\-encoder or SpeechLLM forward pass\.

## 3Experimental Setup

### 3\.1Datasets and Protocol

We train the phoneme\-CTC front\-ends on the 460\-hour clean subset and the full 960\-hour training set of LibriSpeech\[[15](https://arxiv.org/html/2609.28727#bib.bib19)\]\. Evaluation is conducted on test\-clean and test\-other, containing 2,620 and 2,939 utterances, respectively\. Following the Rare5k protocol\[[12](https://arxiv.org/html/2609.28727#bib.bib1),[8](https://arxiv.org/html/2609.28727#bib.bib7)\], each utterance is paired with its oracle rare words andN∈\{100,500,1000,2000\}N\\in\\\{100,500,1000,2000\\\}distractors\. PTC Retrieval selects at most ten words from this list and supplies them as a textual prompt to Prompt\-SLAM\-ASR\-7B\[[20](https://arxiv.org/html/2609.28727#bib.bib5)\]or Qwen3\-ASR\-0\.6B\[[18](https://arxiv.org/html/2609.28727#bib.bib18)\]\. Both SpeechLLM backbones are kept frozen\.

### 3\.2Models and Implementation

We compare DS\-KWS\[[3](https://arxiv.org/html/2609.28727#bib.bib14),[2](https://arxiv.org/html/2609.28727#bib.bib16)\], WavLM\[[5](https://arxiv.org/html/2609.28727#bib.bib20)\], and Qwen3\-ASR’s audio Transformer \(AuT\)\[[18](https://arxiv.org/html/2609.28727#bib.bib18)\]as phoneme front\-ends\. DS\-KWS is a 3\.61M\-parameter KWS baseline operating at 25 Hz\. For WavLM and AuT, we freeze the encoders and train only a layer mixture, a two\-layer Transformer, and a 71\-class phoneme\-CTC head\. WavLM operates at 50 Hz with 317\.31M total and 1\.86M trainable parameters, while AuT operates at 12\.5 Hz with 186\.48M total and 1\.83M trainable parameters\. Training uses WeNet\[[21](https://arxiv.org/html/2609.28727#bib.bib21)\]with a batch size of 8 on four NVIDIA RTX 3090 GPUs\.

PTC Retrieval usesM=100M=100,K=10K=10,β=1\.1\\beta=1\.1,ρ=0\\rho=0,τ=0\.25\\tau=0\.25, andη=0\.05\\eta=0\.05\. PTC Correction extends each local window by five posterior frames on both sides and applies an acoustic\-margin threshold of 2\.0\. Both stages share the cached phoneme posteriors\.

### 3\.3Evaluation Metrics

We report phoneme error rate \(PER\), overall word error rate \(WER\), biased\-word error rate \(B\-WER\), and unbiased\-word error rate \(U\-WER\)\. Following BR\-ASR\[[8](https://arxiv.org/html/2609.28727#bib.bib7)\], retrieval is evaluated usingRecallB​@​99\\mathrm\{Recall\}\_\{B\}@99, the mean number of candidates required to reach 99% oracle\-word recall;RecallB​\#​50\\mathrm\{Recall\}\_\{B\}\\\#50, the oracle\-word recall within the top 50; andRecallH​\#​50\\mathrm\{Recall\}\_\{H\}\\\#50, the near\-homophone distractor recall within the top 50\.

## 4Experimental Results

### 4\.1Main Results

Table[1](https://arxiv.org/html/2609.28727#S2.T1)compares PTC\-Bias with DB\-NNLM\[[12](https://arxiv.org/html/2609.28727#bib.bib1)\], USTR\-CT\[[17](https://arxiv.org/html/2609.28727#bib.bib3)\], CB\-QwenAudio\[[7](https://arxiv.org/html/2609.28727#bib.bib4)\], CTC\-Filter\[[20](https://arxiv.org/html/2609.28727#bib.bib5)\], and Bias Retrieval\[[8](https://arxiv.org/html/2609.28727#bib.bib7)\]\. AtN=2000N=2000, PTC\-Bias with Prompt\-SLAM\-ASR\-7B achieves WERs of 1\.30%/2\.93% and B\-WERs of 3\.38%/7\.63% on test\-clean/test\-other\. Compared with CTC\-Filter, it reduces B\-WER by 23\.4%/23\.9% while keeping U\-WER nearly unchanged\. With Qwen3\-ASR\-0\.6B and its AuT phoneme front\-end\[[18](https://arxiv.org/html/2609.28727#bib.bib18)\], PTC\-Bias achieves WERs of 1\.54%/3\.37% and B\-WERs of 4\.45%/8\.61%\. The consistent gains fromN=100N=100toN=2000N=2000demonstrate its robustness to large bias lists\.

The nested PTC Correction rows further show the benefit of post\-decoding competition\. AtN=2000N=2000, it reduces B\-WER from 3\.85%/8\.24% to 3\.38%/7\.63% with Prompt\-SLAM\-ASR\-7B, and from 5\.08%/9\.24% to 4\.45%/8\.61% with Qwen3\-ASR\. This indicates that retrieving the correct bias word does not always guarantee correct generation, while local competition can further resolve near\-homophone confusions and span\-boundary errors\. The†\\daggerrows reuse the WavLM\[[5](https://arxiv.org/html/2609.28727#bib.bib20)\]retrieval results from Prompt\-SLAM\-ASR for Qwen3\-ASR decoding; under this controlled setting, PTC\-Bias reduces B\-WER by 14\.8%/17\.9% relative to CTC\-Filter\.

### 4\.2Phoneme Modeling and Bias Retrieval

Table[2](https://arxiv.org/html/2609.28727#S4.T2)compares the phoneme front\-ends, using DS\-KWS\[[3](https://arxiv.org/html/2609.28727#bib.bib14),[2](https://arxiv.org/html/2609.28727#bib.bib16)\]as a lightweight KWS\-oriented baseline\. With LS\-960 training, WavLM achieves PERs of 1\.13%/2\.27% on test\-clean/test\-other, compared with 2\.06%/5\.13% for AuT and 4\.45%/11\.80% for DS\-KWS\. In particular, the AuT results indicate that existing SpeechLLM audio encoders provide sufficient phoneme discrimination to support both PTC stages\.

Table[3](https://arxiv.org/html/2609.28727#S4.T3)further compares PTC Retrieval with BR\-ASR\[[8](https://arxiv.org/html/2609.28727#bib.bib7)\]\. Relative to acoustic BR\-ASR, PTC Retrieval reducesRecallB​@​99\\mathrm\{Recall\}\_\{B\}@99from 42\.2 to 16\.9 andRecallH​\#​50\\mathrm\{Recall\}\_\{H\}\\\#50from 69\.3% to 22\.0%, while maintainingRecallB​\#​50\\mathrm\{Recall\}\_\{B\}\\\#50at 99\.3%\. PTC Retrieval therefore reaches high oracle\-word recall with a smaller shortlist and substantially reduces near\-homophone interference\.

Table 2:Phoneme error rates \(PER, %\) of different phoneme front\-ends on LibriSpeech\. Lower is better\.Table 3:Bias retrieval on LibriSpeech test\-other withN=2000N=2000distractors\. Acoustic and Textual are BR\-ASR\[[8](https://arxiv.org/html/2609.28727#bib.bib7)\]variants \.
### 4\.3Qualitative and Efficiency Analysis

Figure[2](https://arxiv.org/html/2609.28727#S4.F2)visualizes phoneme\-level temporal competition in PTC Retrieval\. Similar pronunciations often produce overlapping high\-scoring paths over the same speech segment\. By treating these paths as competing explanations, PTC Retrieval suppresses weaker near\-homophones while retaining confident events at distinct intervals, producing a compact and less ambiguous shortlist\.

Figure[3](https://arxiv.org/html/2609.28727#S4.F3)reports CPU retrieval latency for a 5\.09\-s utterance, excluding phoneme\-posterior computation and trie construction\. AtN=2000N=2000, retrieval takes 7\.61 ms with four workers\. ForN=50,000N=50\{,\}000, increasing the worker count from one to 16 reduces latency from 391\.24 ms to 34\.74 ms\. These results show that CPU parallelism enables PTC Retrieval to scale efficiently to large bias lists\.

![Refer to caption](https://arxiv.org/html/2609.28727v1/tcfsd_n500_competition_mel_4852_tight.png)Figure 2:Temporal competition among phoneme paths in PTC Retrieval\. Solid paths are selected; dashed paths are competing alternatives\.![Refer to caption](https://arxiv.org/html/2609.28727v1/speed.png)Figure 3:CPU search latency of PTC Retrieval for a 5\.09\-s utterance across bias\-list sizes and worker counts\. Phoneme\-posterior extraction and trie construction are excluded\.

## 5Conclusion

We presented PTC\-Bias, a two\-stage contextual biasing framework for SpeechLLMs based on phoneme\-level temporal competition\. Before decoding, PTC Retrieval uses frame\-level phoneme posteriors to select bias words and locate their speech intervals; after decoding, PTC Correction compares misrecognized transcript spans with the retrieved candidates\. Both stages reuse the same phoneme posteriors without an additional SpeechLLM forward pass\. Experiments on LibriSpeech show consistent improvements with 100–2000 distractors, different phoneme front\-ends, and two SpeechLLM backbones\. PTC\-Bias substantially reduces B\-WER while largely preserving U\-WER and maintains efficient CPU retrieval for large bias lists\.

## References

- \[1\]J\. Agrawal, B\. Koilakuntla, and S\. Konjeti\(2025\)Spot and Merge: A Hybrid Context Biasing Approach for Rare Word and Out of Vocabulary Recognition\.InProceedings of Interspeech,pp\. 3319–3323\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2025-2692)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p4.1)\.
- \[2\]Z\. Ai, H\. Cheng, S\. Mu, X\. Li, Y\. Zhou, and S\. Xu\(2026\)Effective User\-Defined Keyword Spotting With Dual\-Stage Matching, Multi\-Modal Enrollment, and Continual Adaptation\.IEEE Transactions on Audio, Speech and Language Processing34,pp\. 3214–3227\.External Links:[Document](https://dx.doi.org/10.1109/TASLPRO.2026.3695299)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p3.1),[§2\.4](https://arxiv.org/html/2609.28727#S2.SS4.p1.1),[§3\.2](https://arxiv.org/html/2609.28727#S3.SS2.p1.1),[§4\.2](https://arxiv.org/html/2609.28727#S4.SS2.p1.1),[Table 2](https://arxiv.org/html/2609.28727#S4.T2.2.3.1.1)\.
- \[3\]Z\. Ai, H\. Cheng, Y\. Wang, S\. Mu, Y\. Zhou, and S\. Xu\(2026\)Dual Data Scaling for Robust Two\-Stage User\-Defined Keyword Spotting\.InICASSP 2026 \- 2026 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 18307–18311\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP55912.2026.11460636)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p3.1),[§3\.2](https://arxiv.org/html/2609.28727#S3.SS2.p1.1),[§4\.2](https://arxiv.org/html/2609.28727#S4.SS2.p1.1),[Table 2](https://arxiv.org/html/2609.28727#S4.T2.2.3.1.1)\.
- \[4\]A\. Andrusenko, A\. Laptev, V\. Bataev, V\. Lavrukhin, and B\. Ginsburg\(2024\)Fast Context\-Biasing for CTC and Transducer ASR Models with CTC\-Based Word Spotter\.InProceedings of Interspeech,pp\. 757–761\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-1002)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p3.1),[§1](https://arxiv.org/html/2609.28727#S1.p4.1),[§2\.4](https://arxiv.org/html/2609.28727#S2.SS4.p1.1)\.
- \[5\]S\. Chen, C\. Wang, Z\. Chen, Y\. Wu, S\. Liu, Z\. Chen, J\. Li, N\. Kanda, T\. Yoshioka, X\. Xiao, J\. Wu, L\. Zhou, S\. Ren, Y\. Qian, Y\. Qian, J\. Wu, M\. Zeng, and F\. Wei\(2022\)WavLM: Large\-Scale Self\-Supervised Pre\-Training for Full Stack Speech Processing\.IEEE Journal of Selected Topics in Signal Processing16\(6\),pp\. 1505–1518\.External Links:[Document](https://dx.doi.org/10.1109/JSTSP.2022.3188113)Cited by:[§3\.2](https://arxiv.org/html/2609.28727#S3.SS2.p1.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.28727#S4.T2.2.5.1.1)\.
- \[6\]N\. Flemotomos, R\. Hsiao, P\. Swietojanski, T\. Hori, D\. Can, and X\. Zhuang\(2025\)Optimizing Contextual Speech Recognition Using Vector Quantization for Efficient Retrieval\.IEEE Transactions on Audio, Speech and Language Processing33,pp\. 3508–3521\.Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p2.1)\.
- \[7\]X\. Gong, A\. Lv, Z\. Wang, and Y\. Qian\(2024\)Contextual Biasing Speech Recognition in Speech\-Enhanced Large Language Model\.InProceedings of Interspeech,pp\. 257–261\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-965)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.5.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p1.1)\.
- \[8\]X\. Gong, A\. Lv, W\. Zhang, Z\. Wang, H\. Zhu, and Y\. Qian\(2025\)BR\-ASR: A Unified Framework for Contextualized Speech Recognition with LLM\-Based Retrieval and Reasoning\.InProceedings of Interspeech,pp\. 4043–4047\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2025-326)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1),[§1](https://arxiv.org/html/2609.28727#S1.p2.1),[§2\.2](https://arxiv.org/html/2609.28727#S2.SS2.p1.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.8.1),[§3\.1](https://arxiv.org/html/2609.28727#S3.SS1.p1.1),[§3\.3](https://arxiv.org/html/2609.28727#S3.SS3.p1.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p1.1),[§4\.2](https://arxiv.org/html/2609.28727#S4.SS2.p2.1),[Table 3](https://arxiv.org/html/2609.28727#S4.T3),[Table 3](https://arxiv.org/html/2609.28727#S4.T3.2.2.1.1),[Table 3](https://arxiv.org/html/2609.28727#S4.T3.2.3.1.1)\.
- \[9\]H\. Hou, X\. Gong, W\. Zhang, W\. Wang, and Y\. Qian\(2025\)Ranking and Selection of Bias Words for Contextual Bias Speech Recognition\.InProceedings of Interspeech,pp\. 5183–5187\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2025-646)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p2.1)\.
- \[10\]L\. Huang, B\. Li, J\. Zhang, L\. Lu, and Z\. Ma\(2023\)Text\-only Domain Adaptation Using Unified Speech\-Text Representation in Transducer\.InProceedings of Interspeech,pp\. 386–390\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2023-1313)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1)\.
- \[11\]Z\. Huang, D\. Caseiro, K\. Joshi, C\. Li, P\. Rondon, Z\. Wu, P\. Zadrazil, and L\. Zhou\(2024\)Optimizing Large\-Scale Context Retrieval for End\-to\-End ASR\.InProceedings of Interspeech,pp\. 4573–4577\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-558)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p2.1)\.
- \[12\]D\. Le, M\. Jain, G\. Keren, S\. Kim, Y\. Shi, J\. Mahadeokar, J\. Chan, Y\. Shangguan, C\. Fuegen, O\. Kalinli, Y\. Saraf, and M\. L\. Seltzer\(2021\)Contextualized Streaming End\-to\-End Speech Recognition with Trie\-Based Deep Biasing and Shallow Fusion\.InProceedings of Interspeech,pp\. 1772–1776\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2021-1566)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.3.1),[§3\.1](https://arxiv.org/html/2609.28727#S3.SS1.p1.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p1.1)\.
- \[13\]Z\. Lei, X\. Na, M\. Xu, E\. Pusateri, C\. V\. Gysel, Y\. Zhang, S\. Han, and Z\. Huang\(2025\)Contextualization of ASR with LLM Using Phonetic Retrieval\-Based Augmentation\.In2025 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP49660.2025.10888744)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1),[§1](https://arxiv.org/html/2609.28727#S1.p2.1)\.
- \[14\]Y\. Nakagome and M\. Hentschel\(2025\)WCTC\-Biasing: Retraining\-Free Contextual Biasing ASR with Wildcard CTC\-Based Keyword Spotting and Inter\-Layer Biasing\.InProceedings of Interspeech,pp\. 5178–5182\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2025-1300)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p3.1),[§2\.4](https://arxiv.org/html/2609.28727#S2.SS4.p1.1)\.
- \[15\]V\. Panayotov, G\. Chen, D\. Povey, and S\. Khudanpur\(2015\)Librispeech: An ASR Corpus Based on Public Domain Audio Books\.In2015 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 5206–5210\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP.2015.7178964)Cited by:[§3\.1](https://arxiv.org/html/2609.28727#S3.SS1.p1.1)\.
- \[16\]K\. Park and J\. Kim\(2019\)g2pE: a simple python module for english grapheme\-to\-phoneme conversion\.GitHub\.Note:[https://github\.com/Kyubyong/g2p](https://github.com/Kyubyong/g2p)Cited by:[§2\.3](https://arxiv.org/html/2609.28727#S2.SS3.p1.2)\.
- \[17\]J\. Qiu, L\. Huang, B\. Li, J\. Zhang, L\. Lu, and Z\. Ma\(2023\)Improving Large\-Scale Deep Biasing With Phoneme Features and Text\-Only Data in Streaming Transducer\.In2023 IEEE Automatic Speech Recognition and Understanding Workshop \(ASRU\),pp\. 1–8\.External Links:[Document](https://dx.doi.org/10.1109/ASRU57964.2023.10389716)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.4.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p1.1)\.
- \[18\]Qwen Team\(2026\)Qwen3\-ASR Technical Report\.arXiv preprint arXiv:2601\.21337\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2601.21337)Cited by:[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.11.1),[§3\.1](https://arxiv.org/html/2609.28727#S3.SS1.p1.1),[§3\.2](https://arxiv.org/html/2609.28727#S3.SS2.p1.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.28727#S4.T2.2.4.1.1)\.
- \[19\]Y\. Xi, H\. Li, X\. Gu, Y\. Jiang, and K\. Yu\(2025\)MFA\-KWS: Effective Keyword Spotting With Multi\-Head Frame\-Asynchronous Decoding\.IEEE Transactions on Audio, Speech and Language Processing33,pp\. 3014–3027\.External Links:[Document](https://dx.doi.org/10.1109/TASLPRO.2025.3587459)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p3.1),[§2\.4](https://arxiv.org/html/2609.28727#S2.SS4.p1.1)\.
- \[20\]G\. Yang, Z\. Ma, Z\. Gao, S\. Zhang, and X\. Chen\(2024\)CTC\-Assisted LLM\-Based Contextual ASR\.In2024 IEEE Spoken Language Technology Workshop \(SLT\),pp\. 126–131\.External Links:[Document](https://dx.doi.org/10.1109/SLT61566.2024.10832154)Cited by:[§1](https://arxiv.org/html/2609.28727#S1.p1.1),[§1](https://arxiv.org/html/2609.28727#S1.p2.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.14.1.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.6.1),[Table 1](https://arxiv.org/html/2609.28727#S2.T1.2.1.7.1),[§3\.1](https://arxiv.org/html/2609.28727#S3.SS1.p1.1),[§4\.1](https://arxiv.org/html/2609.28727#S4.SS1.p1.1)\.
- \[21\]Z\. Yao, D\. Wu, X\. Wang, B\. Zhang, F\. Yu, C\. Yang, Z\. Peng, X\. Chen, L\. Xie, and X\. Lei\(2021\)WeNet: Production Oriented Streaming and Non\-Streaming End\-to\-End Speech Recognition Toolkit\.InProceedings of Interspeech,pp\. 4054–4058\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2021-1983)Cited by:[§3\.2](https://arxiv.org/html/2609.28727#S3.SS2.p1.1)\.