Paper deep dive
Topic Matching in the Wild: Benchmark and Lessons from Real-World ASR Transcripts
Saman Rahbar, Xiliang Zhu, Irvin Cardoza, David Rossouw
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/2/2026, 3:42:17 AM
Summary
This paper benchmarks topic matching in real-world contact center environments using noisy Automatic Speech Recognition (ASR) transcripts. The authors curate a human-annotated dataset of 2,655 utterance-topic pairs and compare three matcher types: regex baselines, zero-shot sentence embedding encoders, and Large Language Model (LLM) matchers (specifically Gemini variants). They evaluate two topic representations: keyphrase lists and natural language descriptions. Results indicate that lightweight LLMs (Gemini-3-Flash and Gemini-2.5-Flash-Lite) paired with natural language descriptions achieve the highest F1 score (0.847), significantly outperforming regex and embedding models. The study highlights that topic representation is as critical as the matcher choice, with LLMs favoring descriptions and embeddings favoring keyphrases.
Entities (10)
Relation Signals (7)
Gemini 3 Flash â achievesbestperformanceon â Natural Language Description
confidence 95% · The strongest matcher is an LLM reading a natural-language description. Gemini-3-Flash reaches F1=0.847
ASR Transcripts â contains â Disfluencies
confidence 95% · These are noisy user-generated text: disfluencies and repetitions, false starts, residual recognition errors
Gemini 2.5 Flash-Lite â iscosteffectivealternativeto â Gemini 2.5 Pro
confidence 95% · Gemini-2.5-Flash-Lite sits at the opposite corner... ~500x cheaper than 2.5-Pro, at F1 within 0.02 of the best matcher.
Gemini 3 Flash â outperforms â Zero-shot Sentence Embedding Encoders
confidence 95% · The LLM matchers beat the regex baseline... and every zero-shot encoder
Gemini 3 Flash â outperforms â Regex Baseline
confidence 95% · The margin over regex is large and highly significant... ahead of the regex baseline (0.721) by 12.6 points
Zero-shot Sentence Embedding Encoders â performsbeston â Keyphrase List
confidence 95% · embeddings are strongest on keyphrases (0.708 vs. 0.660)
Contact Centers â utilizes â Agent-assist tools
confidence 95% · In contact centers, real-time agent-assist tools determine... whether a live customer utterance is relevant
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In contact centers, real-time agent-assist tools determine, for each of many predefined topics, whether a live customer utterance is relevant and display a coaching card to the agent when it is. The input is noisy and challenging: ASR(Automatic Speech Recognition) transcripts of spontaneous phone conversations, which can be unclear, repetitive, and mostly lack punctuation. To systematically study this real-world task, we curate a human-annotated topic-utterance judgments dataset sourced from real call-center transcripts. We compare three types of matchers: a regex-based baseline, zero-shot sentence embedding encoders, and Gemini-based LLM matchers. In addition, two types of topic representations are studied in our benchmark:keyphrases and natural language description. Our empirical experiments highlight the superior performance of lightweight LLM matchers over embedding and regex models when equipped with natural language descriptions.
Tags
Links
- Source: https://arxiv.org/abs/2609.00330v1
- Canonical: https://arxiv.org/abs/2609.00330v1
Trouble viewing inline? Open PDF directly â
Full Text
38,584 characters extracted from source content.
Expand or collapse full text
Topic Matching in the Wild: Benchmark and Lessons from Real-World ASR Transcripts Saman Rahbar Xiliang Zhu Irvin Cardoza David Rossouw Affiliation: Dialpad Inc. Affiliation: sam.rahbar, xzhu, irvin.cardoza, davidr@dialpad.com Abstract In contact centers, real-time agent-assist tools determine, for each of many predefined topics, whether a live customer utterance is relevant and display a coaching card to the agent when it is. The input is noisy and challenging: ASR (Automatic Speech Recognition) transcripts of spontaneous phone conversations, which can be unclear, repetitive, and mostly lack punctuation. To systematically study this real-world task, we curate a human-annotated topic-utterance judgments dataset sourced from real call-center transcripts. We compare three types of matchers: a regex-based baseline, zero-shot sentence-embedding encoders, and Gemini-based LLM matchers. In addition, two types of topic representations are studied in our benchmark: keyphrases and natural language description. Our empirical experiments highlight the superior performance of lightweight LLM matchers over embedding and regex models when equipped with natural language descriptions. 1 Introduction Noisy utterance u âso can I, can I just get my money back?â Keyphrase list refund, money back, ⊠Description âwants their money backâ Topic t (two representations) Regex (baseline) /refund/ /cancel/ ⊠Embedding cosâĄ(uâ,tâ)>Ï ( u, t)>Ï LLM â âIs u about t?â â YES/NO Coaching card to the agent u Figure 1: Our evaluation space. A live call yields a noisy utterance u; a predefined topic t is expressed two ways: a keyphrase list or a natural-language description. Every matcher scores the utterance against a topic representation: a regex baseline (keyphrases only), a zero-shot embedding (cosine similarity), or an LLM prompted for a YES/NO decision. A positive match surfaces a coaching card to the agent. We evaluate every applicable (utterance, representation) pairing; the best (â , blue) is an LLM reading a description. Contact centers can use real-time âagent-assistâ software to support human agents during a live call. Streaming Automatic Speech Recognition (ASR) transcribes the customer, and based on a curated list of topics, the system decides whether any of them is present in the conversation. When a topic matches the current utterance, the system displays a coaching card for the agent (Fig. 1). This requires low-latency processing of hundreds of topics simultaneously (Rawat and Barres, 2022; Barrionuevo-Valenzuela et al., 2026). Additionally, the input text is a challenging ASR transcript of spontaneous speech containing disfluencies (e.g. âmy, my, my, the, my vehicleâ), false starts, and residual recognition errors. These affect downstream processing pipelines (Ritter et al., 2011), and the errors are non-negligible (Feng et al., 2021; Shapira et al., 2025). The standard approach is to clean the text, for example, with lexical normalization (Han and Baldwin, 2011; van der Goot et al., 2021) or disfluency removal (Honnibal and Johnson, 2014; Zayats et al., 2016). We consider the binary topicâutterance matching problem: given an utterance recognized with ASR, and a topic definition, decide whether the utterance matches the topic. To avoid additional latency, we process the raw ASR transcript. We compare two representations for the topic definition: the keyphrase lists that can be compiled into a regex matching program, and natural-language descriptions that can be processed by more advanced models. Both are evaluated on a human-annotated internal dataset. Various methods are benchmarked for matching utterances to topics: the regex baseline, zero-shot embedding similarity (Reimers and Gurevych, 2019), and Gemini LLM matchers (Gemini Team, 2025). We summarize our empirical contributions: 1. C1: to our knowledge, the first study of topic matching on real, noisy production ASR, with the noise measured. Previous resources are social-media text (Baldwin et al., 2015; Derczynski et al., 2017; van der Goot et al., 2021), read or synthetic spoken-language understanding (Bastianelli et al., 2020; Shon et al., 2022; Feng et al., 2021), or clean crowdsourced intent sets (Larson et al., 2019; Coucke et al., 2018; Casanueva et al., 2020). None target the noisy textual input setting. We measure the noise directly (utterance length, 11.4%11.4\% disfluent repetitions, 24.2%24.2\% filler markers). However, our transcripts are sensitive customer data, so we release the evaluation protocol and findings rather than the data. 2. C2: an efficient Flash-tier LLM is enough. The LLM matchers beat the regex baseline (the best by 12.612.6 F1F_1 points, p<10â14p<10^-14) and every zero-shot encoder, even though encoders are the standard tool for semantic matching. The best two matchers are Flash-tier (Gemini-3-Flash and Gemini-2.5-Flash-Lite), and the heavier (e.g. Gemini-2.5-Pro) does not beat them. So a small, fast model is the practical choice for a latency-bounded deployment. 3. C3: the best way to represent and match a topic. A topic can be written as a keyphrase list or as a natural-language description; we test every matcher against both and report the optimal combination. The strongest is an LLM reading a natural-language description (F1=0.847F_1=0.847), and the best representation differs by methodâLLMs favor a description, embeddings a keyphrase list. How a topic is represented is thus as decisive as which matcher scores it. 2 Related Work Noisy user-generated text and normalization. Applying standard NLP to non-canonical text is a long-standing problem. Non-standard vocabulary and syntax break tools trained on edited corpora (Eisenstein, 2013), and off-the-shelf pipelines degrade sharply on tweets (Ritter et al., 2011). The classic response is lexical normalization to a canonical vocabulary (Han and Baldwin, 2011), systems like MoNoise built on this idea (van der Goot and van Noord, 2017). Later work took normalization multilingual (van der Goot et al., 2021). For spontaneous speech, the counterpart to orthographic normalization is disfluency detection and removal: finding repetitions, false starts, and self-corrections in transcribed speech (Honnibal and Johnson, 2014; Zayats et al., 2016). These are the phenomena our noise statistics quantify. Our setting is spontaneous customer speech through call-center ASR, not social-media orthography. Rather than normalizing first, we match topics directly to achieve optimal latency performance, and leave a normalize-then-match comparison to future work. ASR-error robustness and SLU. A parallel line studies how transcription errors propagate downstream. Spoken language understanding (SLU) benchmarks push toward natural speech (Bastianelli et al., 2020; Shon et al., 2022). ASR-GLUE isolates natural language understanding (NLU) robustness under transcription noise (Feng et al., 2021). Models trained on clean text degrade on ASR output, and recovering that performance takes deliberate effort (Cui et al., 2021; Huang and Chen, 2020; Shapira et al., 2025). We differ in using real contact-center ASR, not read, with binary topic matching as the downstream task. Zero-shot classification and semantic matching. Zero-shot prompting (Brown et al., 2020) with instruction tuning (Wei et al., 2022) underpins the LLM-based paradigm, in which strong models proxy human judgment (Zheng et al., 2023; Liu et al., 2023; Gilardi et al., 2023). As a non-generative alternative we use zero-shot encoders (Reimers and Gurevych, 2019). Our set spans E5, BGE, MiniLM, and MPNet (Wang et al., 2022; Xiao et al., 2024; Wang et al., 2020; Song et al., 2020), chosen for their Massive Text Embedding Benchmark (MTEB) rankings (Muennighoff et al., 2023). To our best knowledge, no prior work systematically compares LLM matchers, zero-shot embeddings, and a regex baseline for semantic matching over noisy ASR. Contact-center and intent-detection NLP. Our task is short-utterance intent detection. Canonical benchmarks define the family: CLINC150/OOS (Larson et al., 2019), Snips (Coucke et al., 2018), Banking77 (Casanueva et al., 2020), and HWU64 (Liu et al., 2019), alongside short-text classification (Li et al., 2020). These use clean crowdsourced text. Closer to deployment, prior work classifies intent from call-center transcripts (Zhong and Li, 2019) and triggers agent-assist over live support ASR (Rawat and Barres, 2022; Barrionuevo-Valenzuela et al., 2026). Related resources cover policy-driven dialogue (Chen et al., 2021) and multi-annotator customer service (Feigenblat et al., 2021). Our work combines these threads: an LLM matcher on noisy contact-center ASR for real-time agent assist, measured against a regex baseline and zero-shot embeddings. 3 Task and Data 3.1 Task formulation We study binary topicâutterance matching for real-time contact center agent assist. During a live call, streaming ASR transcribes the customer, and for each predefined topic the system decides whether the current utterance11 1 By utterance we mean a single finalized piece of ASR outputâone stretch of the customerâs speech the recognizer has settled on, rather than a whole speaker turn or the entire call. We match each one on its own. is about that topic. A positive decision surfaces a coaching card to the agent. This runs concurrently for many topics per call and is latency-sensitive (Rawat and Barres, 2022; Barrionuevo-Valenzuela et al., 2026). Formally, a matcher fâĄ(u,dâĄ(t))â0,1f(u,d(t))\!â\!\0,1\ maps an utterance u and a topic definition dâĄ(t)d(t) to fire or no-fire; soft-output models produce a score thresholded at Ï. We score against human gold as binary classification, and because the classes are imbalanced (Section 3.2) we foreground precision, recall, and F1F_1 over accuracy. The task is a form of short-utterance intent classification (Li et al., 2020; Larson et al., 2019; Casanueva et al., 2020; Coucke et al., 2018), but on unconstrained, noisy ASR. Two topic representations. A topic can be represented two ways. The first is a keyphrase list: a curated set of trigger phrases based on filler-tolerant regexes. The second is a natural-language description: a short admin-authored sentence of when the topic applies, scored YES/NO by a model . This lets us ask which representation, paired with which matcher, works best, and whether free-text descriptions can replace hand-curated keyphrases. Example. For a topic like requesting a refund, the keyphrase list holds triggers such as refund, money back, and reimburse, while the natural language description reads âThe customer is asking to be refunded or to get their money back.â An utterance like âso can I, can I just get my money back for that?â should match; âmy address is forty-two oak streetâ should not. (Examples are illustrative, not drawn from our real customer data.) 3.2 Data Source. Our benchmark comes from real English customer-service calls at 11 companies from various industries, covering 173 topics/coaching cards. Across those cards the keyphrase lists hold 3,658 keyphrases (2,447 distinct surface forms), and the keyphrase regex runs over the full set rather than a truncated sample. Utterances are noisy ASR. Inputs are ASR transcripts of spontaneous customer speech. This is noisy user-generated text: disfluencies and repetitions, false starts, residual recognition errors, and no reliable sentence structure. These phenomena break off-the-shelf pipelines (Ritter et al., 2011; Han and Baldwin, 2011) and propagate into downstream understanding (Bastianelli et al., 2020; Feng et al., 2021; Shapira et al., 2025; Cui et al., 2021). Noise characterization. Aggregate statistics over the utterances in the samples demonstrate the ânoisyâ nature of our data. The utterances are short and variable: mean 15.815.8 tokens, median 1010, and 11.5%11.5\% have †3 tokens. An immediate word repetition (a stuttered restart) appears in 11.4%11.4\%, and 24.2%24.2\% contain a filler or discourse marker (uh, like, you know; 2.5%2.5\% of all tokens). These statistics quantify spontaneous-speech disfluency. True ASR recognition errors are also present, our ASR system presents a word error rate (WER) of 14.26% for English. 3.3 Annotation and gold standard Every (utterance, topic) pair is labeled independently by three internal annotators, who choose match, no-match, or ambiguous (guidelines in Appendix B). The third option is deliberate: a short, disfluent utterance is often genuinely underspecified, and forcing a binary choice would push that uncertainty into the gold as noise. The gold label is the majority vote, and a pair whose majority is ambiguous, or that has no majority, is set aside rather than guessed (73 pairs: 45 ambiguous, 28 without a majority). Of our annotated pairs, 2,655 carry a decisive positive or negative gold (⌠24% positive) and form the evaluation set, which indicates the imbalance that motivates precision, recall, and F1F_1 over accuracy. On the 2,655 fully voted pairs, inter-annotator agreement is Fleissâ Îș=0.660Îș=0.660 across the three categories, which translates to âsubstantialâ on the LandisâKoch scale (Fleiss, 1971; Landis and Koch, 1977). It sits below perfect for a real reason: deciding whether a half-finished, disfluent utterance is about a topic is a judgment call that is hard for people, not only for models (Derczynski et al., 2017). Together with authentic call center ASR, rather than the social media text (Baldwin et al., 2015; van der Goot et al., 2021) or the read and synthetic speech of SLU benchmarks (Bastianelli et al., 2020; Feng et al., 2021), our consensus-based, multi-annotator gold standard is what defines the new evaluation setting we contribute. 4 Methods We evaluate three matcher families on one shared benchmark, the same 2,6552,655-row human gold set from Section 3.2. The matcher families are (i) a regex baseline, (i) zero-shot sentence-embedding encoders, and (i) instruction-tuned large language model (LLM) matchers. Figure 1 summarizes each family, the topic representation it uses, and how it decides. 4.1 Regex baseline Our baseline defines each topic by its keyphrase list and fires when an utterance matches any keyphrase under a word-boundary regex. The regex tolerates filler âslopâ and normalizes hyphenation, punctuation, and numeric variants. We run it over the full keyphrase set, which contains 3,6583,658 keyphrases, 2,4472,447 distinct, across 173173 coaching cards, so the baseline reflects the keyphrase matcher as actually configured. 4.2 Zero-shot sentence-embedding encoders We embed the utterance and the topic, then score their cosine similarity, following the Sentence-BERT encoder tradition (Reimers and Gurevych, 2019). We use six widely adopted encoders zero-shot, with no fine-tuning: e5-small/base-v2 (Wang et al., 2022), bge-small/base-en-v1.5 (Xiao et al., 2024), all-MiniLM-L6-v2 (Wang et al., 2020), and all-mpnet-base-v2 (Song et al., 2020). We selected them by their standing on the Massive Text Embedding Benchmark (MTEB) (Muennighoff et al., 2023). We run two modes. The description mode takes one cosine to the natural-language description. The keyphrase max-similarity mode takes the max cosine over the topicâs keyphrases. Each mode produces a continuous score that needs a threshold. Tuning that threshold on the test rows would bias the result, so we use 55-fold stratified cross-validation (CV) (Kohavi, 1995). We select the max-F1F_1 threshold on four folds, apply it to the held-out fold, and aggregate. The stratification preserves the ⌠24% positive rate in each fold. 4.3 LLM matchers We recast matching as a prompted YES/NO decision for LLMs. Instruction tuning makes this viable zero-shot (Brown et al., 2020; Wei et al., 2022; Gilardi et al., 2023). We evaluate four Gemini matchers: Gemini-2.5-Flash-Lite, Gemini-2.5-Flash, Gemini-2.5-Pro, and Gemini-3-Flash (preview) (Gemini Team, 2023; Gemini Team, 2024; Gemini Team, 2025)âwhich together cover a range of speed and capability, the trade-off a real-time budget forces. All four use the same prompt (Appendix A) at temperature 00, and each emits a hard YES/NO that we use directly. A few responses come back unparseable (†9 per Gemini-2.5 matcher, none for Gemini-3), and we treat these as no-match (fail-closed). We hold the prompt, decoding, and parsing fixed across all four isolates the model as the source of any performance gap. 4.4 Metrics The classes are imbalanced. Following work on skewed evaluation (Saito and Rehmsmeier, 2015; Davis and Goadrich, 2006; Sokolova and Lapalme, 2009), we center precision, recall, and F1F_1 over accuracy. We score hard-output methods at their native YES/NO decision and embeddings at the CV threshold. We add several agreement and significance measures: 95%95\% bootstrap confidence intervals (CIs) on F1F_1; McNemar tests for the headline contrasts; and F1F_1 on the fully adjudicated subset. 4.5 Cost and latency For each matcher we measure per-decision latency and cost on 200200 topicâutterance pairs sampled from the benchmark, discarding a short warm-up. Latency is the end-to-end time for one (utterance,topic)(utterance,topic) decision under identical conditions: for the LLM matchers, the round trip to the Google Gemini API (temperature 00, one YES/NO token); for embeddings, the utterance encode; for regex, the pattern match. We report it as a relative comparison, not a production service-level objective. We report the marginal API cost per 1,0001,000 decisions for the LLM matchers, from the published per-token prices22 2 Gemini API price list, accessed 2026-08-27. applied to the measured input and output tokens (output includes billed thinking tokens). The underlying compute cost of serving any matcher (instance, CPU/GPU, memory) varies by deployment and is out of scope; we assume a comparable instance-serving cost across all approaches and report only the differential API cost, which the LLM matchers alone incur. The figures are therefore the additional per-decision API cost on top of a common serving baseline, not a total cost of ownership. 5 Results and Discussion Figure 2 gives the headline: the best F1F_1 each matcher reaches. Tables 1 and 2 break the picture out by topic representation, so every matcherârepresentation pairing is visible. Because the classes are imbalanced (⌠24% positive), we report precision, recall, and F1F_1; embeddings are scored at a cross-validated threshold and the other matchers at their native YES/NO decision, and unparseable LLM replies (at most nine per model) count as no-match. 000.20.20.40.40.60.60.80.8Best embeddingRegex2.5 Flash2.5 Pro2.5 Flash-LiteGemini 3 Flash0.8470.8330.8180.8010.7210.708F1F_1 (error bars: 95% bootstrap CI) Figure 2: Main results: F1F_1 on the human-gold benchmark. Error bars are 95%95\% bootstrap CIs for the native-decision methods; the best embedding bar is a 5-fold cross-validated point estimate (fold Ï=0.027Ï\!=\!0.027; Table 1) and is shown without a bootstrap bar as it is not a native-decision method. LLM matchers (blue) lead the regex baseline (orange) and the best of six zero-shot embedding encoders (green); the two lightweight Flash-tier matchers top the board and the LLMâregex gap is significant (p<10â6p<10^-6 for all four matchers). 5.1 An LLM on a description wins The strongest matcher is an LLM reading a natural-language description. Gemini-3-Flash reaches F1=0.847F_1=0.847, ahead of the regex baseline (0.7210.721) by 12.612.6 points and of the best embedding (0.7080.708). The margin over regex is large and highly significant (exact McNemar p<10â6p<10^-6 for every LLM, and p<10â14p<10^-14 for the two best), and it is almost all recall: the LLM catches paraphrases and disfluent phrasings that a fixed keyphrase list misses. Surprisingly, embeddings come last though being the usual backbone of semantic matching in many other tasks. Keyphrases P R F1F_1 Regex (baseline) 0.850 0.625 0.721 all-MiniLM-L6-v2 0.689 0.731 0.708 e5-base-v2 0.673 0.743 0.706 bge-base-en-v1.5 0.669 0.748 0.706 bge-small-en-v1.5 0.657 0.762 0.705 mpnet-base-v2 0.750 0.637 0.689 e5-small-v2 0.620 0.746 0.677 Gemini 3 Flash 0.839 0.752 0.793 2.5 Pro 0.845 0.725 0.780 2.5 Flash-Lite 0.882 0.638 0.740 2.5 Flash 0.898 0.603 0.721 Table 1: All matchers on the keyphrase representation. Embeddings are 5-fold cross-validated; regex and LLMs use their native decision. Best F1F_1 in bold. Natural-language description P R F1F_1 e5-base-v2 0.687 0.635 0.660 all-MiniLM-L6-v2 0.660 0.640 0.649 bge-small-en-v1.5 0.646 0.640 0.641 bge-base-en-v1.5 0.615 0.657 0.635 e5-small-v2 0.646 0.616 0.627 mpnet-base-v2 0.604 0.626 0.613 Gemini 3 Flash 0.866 0.829 0.847 2.5 Flash-Lite 0.887 0.786 0.833 2.5 Pro 0.827 0.809 0.818 2.5 Flash 0.931 0.703 0.801 Table 2: All matchers on the natural-language description representation. Regex is omitted as it matches keyphrase lists only. Embeddings are 5-fold cross-validated. Best F1F_1 in bold. Representation matters as much as the matcher. Tables 1 and 2 show that the best representation is not the same for every method. LLMs are strongest on the description (Gemini-3-Flash 0.8470.847 vs. 0.7930.793 on keyphrases; Gemini-2.5-Flash-Lite 0.8330.833 vs. 0.7400.740); embeddings are strongest on keyphrases (0.7080.708 vs. 0.6600.660); regex applies only to keyphrases, since a description is not a list of patterns. The optimal pairing is therefore an LLM with a natural language description, and how a topic is represented is as consequential as which matcher scores it. A lighter model is enough. The two best systems are the two smallest LLMs, Gemini-3-Flash and Gemini-2.5-Flash-Lite, and no larger model beats them. Gemini-2.5-Flash-Lite is even nominally ahead of the heavier Gemini-2.5-Pro and Gemini-2.5-Flash (p=0.014p=0.014/0.0190.019). This interesting finding suggests that a small Flash-tier model is enough for this task while heavier, most costly LLMs offer no obvious benefits. Cost and latency. Table 3 prices the âlighter is enoughâ finding. Gemini-2.5-Pro is Pareto-dominated: its mandatory thinking (âŒ525 525 output tokens on a YES/NO question) makes it the slowest (p50 5.65.6 s) and by far the costliest ($5.385.38 per 1,0001,000 decisions), yet it does not lead on F1F_1. Gemini-2.5-Flash-Lite sits at the opposite corner: p50 0.350.35 s and $0.0100.010 per 1,0001,000, i.e. âŒ16Ă 16Ă faster and âŒ500Ă 500Ă cheaper than 2.5-Pro, at F1F_1 within 0.020.02 of the best matcher. Embeddings and regex are faster still and self-hosted, but trail on F1F_1 (Tables 1, 2). For a latency- and cost-bounded real-time deployment, a Flash-tier LLM on a description is the practical operating point. Matcher Latency (ms) Tokens $/1k p50 p95 in out Gemini 3 Flash 1742 5018 100 165 0.55 2.5 Flash-Lite 346 528 100 1 0.010 2.5 Pro 5580 8365 100 525 5.38 2.5 Flash 461 771 100 1 0.033 e5-base-v2 29 32 â â â all-MiniLM-L6-v2 6 7 â â â Regex 0 0 â â â Table 3: Per-decision latency and API cost on 200 sampled topicâutterance pairs. LLMs score the natural-language description; embedding latency is the per-utterance encode and regex the keyphrase match. Gemini $/1k is the marginal API cost (prices accessed 2026-08-27; output includes billed thinking tokens). Embeddings and regex are self-hosted and incur no API cost. Latency is measured under identical conditions from one client to the Google Gemini API; embeddings and regex are timed locally on a single CPU core. 5.2 Across three matchers Our central finding is that LLM matchers score significantly higher on this noisy benchmark. What is clear mechanistically is that regex matches only literal keyphrases and cannot absorb disfluencies or recognition errors, and that cosine similarity rewards surface overlap rather than topic semantics when processing noisy textual input as in our setting. 6 Conclusion We studied topicâutterance matching on noisy call-center ASR transcripts, comparing a regex baseline, zero-shot embeddings, and Gemini LLM matchers across two ways of describing a topic. The clear winner is an LLM reading a natural-language description: Gemini-3-Flash reaches an F1F_1 of 0.8470.847, well above the regex baseline (0.7210.721) and the best embedding (â0.71â 0.71). Our comprehensive benchmark demonstrates that lightweight LLMs are enough for this task. The smallest Flash-tier matchers lead, while larger ones do not present superior performance. Additionally, we prove that the topic representation matters as much as the matcher: LLMs are strongest with a written description, while embeddings work better with a keyphrase list. Because the calls are sensitive customer data, we present this evaluation protocol and findings rather than the data itself. We hope this work provides insights to other practitioners working under a similar setting. Limitations Scope. All data are English and from a single domain, contact-center agent assist (11 companies, 173 coaching cards); we make no cross-lingual or cross-domain claims. This is an accuracy result: without a controlled clean-versus-noisy comparison, we do not isolate noise robustness from general semantic capability. Models. Our best matchers are proprietary Gemini API models (Gemini Team, 2025), and the very best (Gemini-3-Flash) is a preview model whose numbers are a snapshot. Our conclusions do not hinge on it: the generally available Gemini-2.5-Flash-Lite already beats every non-LLM baseline, and we anchor the comparison to a fully reproducible regex baseline and open-weight embeddings (Reimers and Gurevych, 2019). We do not test other LLM API services beyond Gemini due to our data safety policy. Data availability. The underlying transcripts contain sensitive customer information and cannot be released. To support reproducibility within this constraint, we release the full evaluation protocol, prompts, matcher configurations, and aggregate statistics, and we report results only in aggregate; access to the data may be considered on a controlled basis subject to the serviceâs data-processing agreements. Ethics and Broader Impact The data is derived from customerâagent dialogues about sensitive topics and processed in accordance with the serviceâs data-processing and consent procedures; we use only the transcript text and do not seek to re-identify any customers or agents. We report only aggregated results, model performance scores, and evaluation methodology, not the data itself. Gold labels were obtained from internal specialized annotators who were compensated fairly. References Baldwin et al. (2015) Timothy Baldwin, Marie Catherine de Marneffe, Bo Han, Young-Bum Kim, Alan Ritter, and Wei Xu. 2015. Shared tasks of the 2015 workshop on noisy user-generated text: Twitter lexical normalization and named entity recognition. In Proceedings of the Workshop on Noisy User-generated Text (WNUT 2015), ACL-IJCNLP. Barrionuevo-Valenzuela et al. (2026) Juan Barrionuevo-Valenzuela, Daniel CalderĂłn-GonzĂĄlez, Zoraida Callejas, and David Griol. 2026. Supporting human operators during customer service interactions with agentic-rag. In Proceedings of the 16th International Workshop on Spoken Dialogue System Technology (IWSDS 2026). Bastianelli et al. (2020) Emanuele Bastianelli, Andrea Vanzo, Pawel Swietojanski, and Verena Rieser. 2020. Slurp: A spoken language understanding resource package. arXiv preprint arXiv:2011.13205. Brown et al. (2020) Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, and 12 others. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Casanueva et al. (2020) Iñigo Casanueva, Tadas TemÄinas, Daniela Gerz, Matthew Henderson, and Ivan VuliÄ. 2020. Efficient intent detection with dual sentence encoders. arXiv preprint arXiv:2003.04807. Chen et al. (2021) Derek Chen, Howard Chen, Yi Yang, Alexander Lin, and Zhou Yu. 2021. Action-based conversations dataset: A corpus for building more in-depth task-oriented dialogue systems. arXiv preprint arXiv:2104.00783. Coucke et al. (2018) Alice Coucke, Alaa Saade, Adrien Ball, ThĂ©odore Bluche, Alexandre Caulier, David Leroy, ClĂ©ment Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, MaĂ«l Primet, and Joseph Dureau. 2018. Snips voice platform: an embedded spoken language understanding system for private-by-design voice interfaces. arXiv preprint arXiv:1805.10190. Cui et al. (2021) Tong Cui, Jinghui Xiao, Liangyou Li, Xin Jiang, and Qun Liu. 2021. An approach to improve robustness of nlp systems against asr errors. arXiv preprint arXiv:2103.13610. Davis and Goadrich (2006) Jesse Davis and Mark Goadrich. 2006. The relationship between precision-recall and roc curves. In Proceedings of the 23rd International Conference on Machine Learning (ICML 2006), p. 233-240. Derczynski et al. (2017) Leon Derczynski, Eric Nichols, Marieke van Erp, and Nut Limsopatham. 2017. Results of the wnut2017 shared task on novel and emerging entity recognition. In Proceedings of the 3rd Workshop on Noisy User-generated Text (W-NUT 2017). Eisenstein (2013) Jacob Eisenstein. 2013. What to do about bad language on the internet. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2013). Feigenblat et al. (2021) Guy Feigenblat, Chulaka Gunasekara, Benjamin Sznajder, Sachindra Joshi, David Konopnicki, and Ranit Aharonov. 2021. Tweetsumm â a dialog summarization dataset for customer service. arXiv preprint arXiv:2111.11894. Feng et al. (2021) Lingyun Feng, Jianwei Yu, Deng Cai, Songxiang Liu, Haitao Zheng, and Yan Wang. 2021. Asr-glue: A new multi-task benchmark for asr-robust natural language understanding. arXiv preprint arXiv:2108.13048. Fleiss (1971) Joseph L. Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological Bulletin 76(5):378-382. Gemini Team (2023) Gemini Team. 2023. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805. Gemini Team (2024) Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530. Gemini Team (2025) Gemini Team. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261. Gilardi et al. (2023) Fabrizio Gilardi, Meysam Alizadeh, and MaĂ«l Kubli. 2023. Chatgpt outperforms crowd workers for text-annotation tasks. arXiv preprint arXiv:2303.15056. Han and Baldwin (2011) Bo Han and Timothy Baldwin. 2011. Lexical normalisation of short text messages: Makn sens a #twitter. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL-HLT 2011). Honnibal and Johnson (2014) Matthew Honnibal and Mark Johnson. 2014. Joint incremental disfluency detection and dependency parsing. Transactions of the Association for Computational Linguistics, 2:131â142. Huang and Chen (2020) Chao-Wei Huang and Yun-Nung Chen. 2020. Learning asr-robust contextualized embeddings for spoken language understanding. arXiv preprint arXiv:1909.10861. Kohavi (1995) Ron Kohavi. 1995. A study of cross-validation and bootstrap for accuracy estimation and model selection. In Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI 1995), Volume 2, p. 1137-1143. Landis and Koch (1977) J. Richard Landis and Gary G. Koch. 1977. The measurement of observer agreement for categorical data. Biometrics 33(1):159-174. Larson et al. (2019) Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars. 2019. An evaluation dataset for intent classification and out-of-scope prediction. arXiv preprint arXiv:1909.02027. Li et al. (2020) Qian Li, Hao Peng, Jianxin Li, Congying Xia, Renyu Yang, Lichao Sun, Philip S. Yu, and Lifang He. 2020. A survey on text classification: From shallow to deep learning. arXiv preprint arXiv:2008.00364. Liu et al. (2019) Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. 2019. Benchmarking natural language understanding services for building conversational agents. arXiv preprint arXiv:1903.05566. Liu et al. (2023) Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634. Muennighoff et al. (2023) Niklas Muennighoff, Nouamane Tazi, LoĂŻc Magne, and Nils Reimers. 2023. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316. Rawat and Barres (2022) Mrinal Rawat and Victor Barres. 2022. Real-time caller intent detection in human-human customer support spoken conversations. arXiv preprint arXiv:2208.06802. Reimers and Gurevych (2019) Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084. Ritter et al. (2011) Alan Ritter, Sam Clark, Mausam, and Oren Etzioni. 2011. Named entity recognition in tweets: An experimental study. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing (EMNLP). Saito and Rehmsmeier (2015) Takaya Saito and Marc Rehmsmeier. 2015. The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PLOS ONE 10(3):e0118432. Shapira et al. (2025) Ori Shapira, Shlomo E. Chazan, and Amir DN Cohen. 2025. Measuring the effect of transcription noise on downstream language understanding tasks. arXiv preprint arXiv:2502.13645. Shon et al. (2022) Suwon Shon, Ankita Pasad, Felix Wu, Pablo Brusco, Yoav Artzi, Karen Livescu, and Kyu J. Han. 2022. Slue: New benchmark tasks for spoken language understanding evaluation on natural speech. arXiv preprint arXiv:2111.10367. Sokolova and Lapalme (2009) Marina Sokolova and Guy Lapalme. 2009. A systematic analysis of performance measures for classification tasks. Information Processing & Management 45(4):427-437. Song et al. (2020) Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. Mpnet: Masked and permuted pre-training for language understanding. arXiv preprint arXiv:2004.09297. van der Goot et al. (2021) Rob van der Goot, Alan Ramponi, Arkaitz Zubiaga, Barbara Plank, Benjamin Muller, Iñaki San Vicente Roncal, Nikola LjubeĆĄiÄ, Ăzlem ĂetinoÄlu, Rahmad Mahendra, Talha ĂolakoÄlu, Timothy Baldwin, Tommaso Caselli, and Wladimir Sidorenko. 2021. Multilexnorm: A shared task on multilingual lexical normalization. In Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT 2021). van der Goot and van Noord (2017) Rob van der Goot and Gertjan van Noord. 2017. Monoise: Modeling noise using a modular normalization system. arXiv preprint arXiv:1710.03476. Wang et al. (2022) Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533. Wang et al. (2020) Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. arXiv preprint arXiv:2002.10957. Wei et al. (2022) Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652. Xiao et al. (2024) Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. arXiv preprint arXiv:2309.07597. Zayats et al. (2016) Vicky Zayats, Mari Ostendorf, and Hannaneh Hajishirzi. 2016. Disfluency detection using a bidirectional LSTM. In Interspeech. Zheng et al. (2023) Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685. Zhong and Li (2019) Junmei Zhong and William Li. 2019. Predicting customer call intent by analyzing phone call transcripts based on cnn for multi-class classification. arXiv preprint arXiv:1907.03715. Appendix A Evaluation Prompt All LLM matchers receive the identical prompt below (temperature 00). The system instruction fixes the task and the strict YES/NO output format; the user template is filled per example with the topic description and the customer utterance. System instruction. You are a precise judge evaluating whether a customer utterance matches a topic description. The topic description defines what a customer might say when they are discussing that topic. You must reply with exactly one word: YES or NO. Do not add any explanation. User template. Topic description: description Customer utterance: âutteranceâ Does the utterance match the topic? Reply YES or NO. Appendix B Annotation Guidelines Each item shows one customer utterance and one topic (its natural-language description). Up to three annotators independently answer a single question, presented verbatim as: Does the utterance fit the description? Would you say the utterance fits under the topic described? with response options yes (positive / match), no (negative / no-match), or ambiguous when the utterance is too underspecified to decide, plus an optional free-text notes field for rationale. Annotators judge semantic fitâwhether the customer is talking about the topicârather than surface keyword overlap, and are instructed to rely only on the utterance itself (no surrounding call context). The per-item gold label is the majority vote over the available annotators. Rows whose majority label is ambiguous, or with no majority (disputed), are excluded from the usable evaluation set; rows with only a single annotator so far are used as per-row gold but flagged and excluded from all inter-annotator agreement statistics (Section 3.2).