PaperlayerView exact revision on arXiv

Source-linked AI summary

When to Call an LLM: A Confidence-Gated Hybrid for Cost-Effective Emotion Recognition in Conversational AI

Sai Babu Udayagiri, Arjun Chouhan, Ravisekhar Kanagala, Trishala Pavagada

arXiv:2609.17977v1cs.AI

TL;DR

Dialogue-contextual emotion recognition must balance accuracy with the cost and latency constraints of CCaaS deployment, and it is unclear whether off-the-shelf LLM prompting beats a lightweight ensemble. The paper compares both systems and a confidence-gated hybrid; the pure-system winner varies by dataset, while the hybrid Pareto-dominates both pure systems across the three evaluated datasets.

  • Problem

    The paper asks whether off-the-shelf LLM prompting outperforms a well-engineered lightweight ensemble on dialogue-contextual ERC, where CCaaS cost and latency constraints also matter.

    Source: §1 · p. 1

  • Method

    The paper compares a lightweight ensemble with GPT-4o-mini prompting and routes the ensemble’s least-confident predictions to an LLM in a confidence-gated hybrid.

    Source: §1 · pp. 1–2

  • Results

    The pure-system winner varies by dataset, while the hybrid Pareto-dominates both pure systems across three datasets and two LLM providers.

    Source: §1 · p. 2, Accuracy: Neither Pure System Is a… · p. 3, §6 · p. 6

  • Takeaways & Limitations

    The hybrid offers a task-specific alternative to choosing either pure system and provides an auditable escalation signal associated with emotion or sentiment shifts.

    Source: §5 · pp. 5–6

    The results have not been piloted on real contact-center transcripts, so transfer from academic benchmarks to that setting remains untested.

    Source: §5 · p. 6

Abstract

from arXiv · show

Emotion recognition in conversation (ERC) is a production capability behind agent-assist prompts, escalation routing, and post-call analytics in contact-center-as-a-service (CCaaS) platforms, where cost and latency constraints matter as much as accuracy. We report a systems-level comparison of three deployment options for dialogue-contextual ERC: a low-cost stacked ensemble (sentence embeddings, windowed context, RandomForest/XGBoost/logistic-regression stacking), off-the-shelf LLM prompting (GPT-4o-mini; zero-shot, few-shot, chain-of-thought), and a confidence-gated hybrid that escalates only the ensemble's least-confident predictions to the LLM - modeled on IVA-to-human-agent escalation policies used in production contact centers. On IEMOCAP, the ensemble significantly outperforms every LLM configuration (0.595 vs. 0.460-0.536 weighted F1, p < 0.0001) at a fraction of the cost and sub-10ms latency; on MELD and CMU-MOSI the ranking reverses, showing neither pure system is a safe default. The confidence-gated hybrid resolves this by Pareto-dominating both pure systems on all three datasets (0.620, 0.643, 0.824 weighted F1) while routing the majority of traffic through the near-zero-cost ensemble, translating to roughly $10-85 per million utterances versus $99-170 for an LLM-only pipeline. The escalation policy is not an opaque cost/accuracy dial: escalated turns disproportionately follow an emotion or sentiment shift, giving operators an interpretable, auditable routing signal, and the ensemble's confidence is well-calibrated and safely under- rather than over-confident. The pattern holds across three datasets and two LLM providers. Confidence-gated cascading is established in general ML systems; our contribution is showing it transfers cleanly to dialogue-contextual ERC, yielding a concrete deployment recipe for CCaaS and conversational-AI platforms deciding how to allocate LLM spend.

1 Introduction

The paper asks whether off-the-shelf LLM prompting outperforms a lightweight ensemble for dialogue-contextual ERC, and whether a hybrid can avoid a dataset-specific deployment bet. It compares these options for CCaaS, where accuracy, cost, and latency all matter.

1 Introduction

  • Production CCaaS applications include agent-assist prompts, automated escalation routing, quality-assurance scoring, and post-call analytics.Latency and throughput constraints motivated prior work on a comparatively simple stacked ensemble for IEMOCAP.Source: §1 · p. 1
  • The deployment choice is three-way: the paper brackets fine-tuned or multimodal ERC architectures and compares options without GPU-serving costs.The ensemble runs on CPU in under 10 ms, while LLM calls incur per-call cost and latency.Source: §1 · p. 1
  • The paper compares a lightweight ensemble, GPT-4o-mini prompting, and a confidence-gated hybrid across three datasets, with paired bootstrap significance testing.The hybrid escalates turns the ensemble is least confident about to an LLM.Source: §1 · pp. 1–2
  • The paper identifies chain-of-thought prompting’s systematic under-prediction of neutral as a mechanism behind its differing performance on IEMOCAP and MELD.The paper also reports that escalation disproportionately follows emotion or sentiment shifts, with a 1.11–1.17× lift, and that ensemble confidence is well-calibrated and mildly under-confident.Source: §1 · p. 2

2 Related Work

Prior ERC work spans context-aware architectures, lightweight ensembles, and LLM prompting, but had not compared a lightweight ensemble with LLM prompting or their confidence-gated hybrid. Existing work on cascading and routing provides related designs, while this paper focuses on calibrated ensemble confidence for dialogue-contextual ERC.

2 Related Work

  • Earlier ERC work includes recurrence-, graph-, and commonsense-based context modeling, followed by fine-tuned transformers and multimodal fusion.Lightweight tree ensembles were reported as competitive with, but not superior to, this lineage at a fraction of the cost.Source: §2 · p. 2
  • Most LLM affect-classification benchmarks use single utterances; Feng et al. (2024) is a rare dialogue-level study and finds LLM prompting generally underperforms fine-tuned neural ERC baselines.Prior work also reports substantial LLM degradation in multi-turn settings (Laban et al. 2025), while chain-of-thought results for affect tasks are mixed (Zheng, Zhao, and Li 2025; Miriyala, Bukkapatnam, and Prahallad 2025).Source: §2 · p. 2
  • Existing cascade designs include LLM-to-LLM routers, non-LLM-to-LLM deferral, and theoretical work on threshold-based deferral.Examples include FrugalGPT (Chen, Zaharia, and Zou 2023), RouteLLM (Ong et al. 2024), and logistic-regression-to-LLM deferral with a learned policy (Nie et al. 2024).Source: §2 · p. 2
  • This paper’s stated distinction is direct gating on a calibrated tree ensemble’s class probability, without a learned deferral policy or LLM-judged quality score.The authors position its transfer to dialogue-contextual ERC, cross-dataset performance, and interpretable escalation signal as a task-specific contribution.Source: §2 · p. 2

3 Method

The method compares a stacked ensemble, GPT-4o-mini prompting, and a confidence-gated hybrid on full test splits. It selects a routing threshold by sweeping the ensemble’s top-class probability threshold and evaluates weighted F1 as the primary metric.

3 Method

  • The ensemble uses utterance and preceding-context sentence embeddings plus a speaker flag, with RandomForest and XGBoost base learners and a logistic-regression meta-learner.The model uses five-fold internal cross-validation to generate out-of-fold base-learner probabilities; it runs on CPU in under 10 ms per utterance.Source: §3 · pp. 2–3
  • The hybrid returns the ensemble prediction when its top-class probability meets threshold τ; otherwise it calls the LLM and falls back to the ensemble if parsing fails.The routing algorithm’s outcomes and fallback behavior are specified directly in the algorithm passage.Source: §3 · p. 3
  • GPT-4o-mini is evaluated with zero-shot, six-shot few-shot, and six-shot chain-of-thought prompts using a five-turn context window and temperature 0.Llama-3-8B-Instruct is also evaluated on the full test splits as an independent second provider.Source: §3 · p. 3
  • The threshold is swept from 0.00 to 1.00 in increments of 0.05 to trace accuracy versus cost against pure-ensemble and pure-LLM endpoints.The reported threshold is each dataset’s empirically best τ; the method notes that this selection uses a fully labeled evaluation split.Source: §3 · p. 3
  • The Table 2 comparison reports weighted F1 for the ensemble and best LLM strategy on full test splits, with the winning pure system flipping by dataset.The caption specifies paired-bootstrap 95% confidence intervals for the weighted-F1 difference, but provides no cell values here.Source: Table 2 · p. 3
  • Weighted F1 is primary and macro F1 secondary; paired bootstrap tests use 1,000 resamples, and the evaluation uses full test splits without subsampling.Table 1 identifies the dataset split provenance: IEMOCAP holds out Session 5, while MELD and CMU-MOSI use their released train/test splits.Source: §3 · p. 3, Table 1 · p. 3

4 Experiments and Results

The pure systems reverse their relative performance across datasets, while confidence-gated routing numerically beats both pure endpoints on all three. Its escalation signal is associated with emotion or sentiment shifts.

Accuracy: Neither Pure System Is a Safe Default

  • IEMOCAP favors the ensemble, MELD zero-/few-shot favors GPT-4o-mini, and CMU-MOSI favors the LLM; these reversals make neither pure system a safe default.On MELD, the zero-/few-shot reversal is specific to GPT-4o-mini; Llama-3-8B loses to the ensemble across all tested strategies.Source: Accuracy: Neither Pure System Is a… · p. 3

Where the Ensemble and LLM Disagree

  • The systems excel on different slices: the LLM wins short-context turns, the ensemble wins long-context turns, and both degrade on emotion-shift turns.On MELD, ensemble minority-class collapse accounts for much of its aggregate deficit; on CMU-MOSI, it never predicts neutral.Source: Where the Ensemble and LLM Disagree · pp. 3–4, Table 3 · p. 4

Cost and Latency in Deployment Terms

  • At one million utterances monthly, the ensemble costs about $10, the recommended hybrid $61–85, and a pure-LLM pipeline $99–170.Ensemble inference is CPU-bound and sub-10ms; LLM calls cost 3.9–17 times more per utterance and add network latency.Source: Cost and Latency in Deployment Terms · p. 4, Table 4 · p. 4

The Confidence-Gated Hybrid Pareto-Dominates Both Endpoints

  • The hybrid numerically beats both pure endpoints on all three datasets; its margin over the better pure system is significant on IEMOCAP and MELD, but not CMU-MOSI.On CMU-MOSI, the hybrid is statistically indistinguishable in accuracy from the pure LLM while costing $85 versus $138 per million utterances.Source: The Confidence-Gated Hybrid… · pp. 4–5, Table 5 · p. 5

The Router Is Interpretable, Not a Black Box

  • Shift turns escalate at a consistent 1.11–1.17× lift over each dataset’s overall rate, but the router is not claimed to explicitly detect emotional dynamics.The ensemble is well-calibrated across datasets (ECE ≤0.046) and mildly under-confident.Source: Table 6 · p. 5, The Router Is Interpretable, Not a… · p. 5

5 Discussion

The discussion frames ERC deployment as task-specific: neither pure system is a safe default, while confidence-gated routing offers a practical alternative. Its transfer to production contact-center data remains untested.

Deployment implications

  • The IEMOCAP-versus-MELD/CMU-MOSI reversal means neither a lightweight ensemble nor off-the-shelf LLM prompting is a universally safe first choice.The hybrid wins on every tested dataset by escalating only the fraction of traffic the ensemble is unsure about, at a fraction of full-LLM cost.Source: §5 · pp. 5–6
  • Escalated turns correlate with emotion or sentiment shifts, offering an operationally interpretable routing signal rather than an opaque probability cutoff.The discussion connects this signal to concepts CCaaS quality teams already use when reasoning about escalation-worthy calls.Source: §5 · p. 6

Implementation and next steps

  • The routing policy needs no LLM fine-tuning or change to existing IVA-to-human handoff infrastructure, only a calibrated confidence score from the classifier already in front of the LLM.The authors suggest testing conversational-state tracking in prompts as a direction for addressing the LLM’s long-context weakness on IEMOCAP.Source: §5 · p. 6

Limitations and open questions

  • The validated cascade is specifically a tree-ensemble-to-LLM recipe for ERC; its generality to other dialogue-structured classification tasks is untested.Replacing the ensemble with a fine-tuned transformer produced mixed rather than uniformly positive results, making the benefit contingent on classifier-specific failure modes.Source: §5 · p. 6
  • Transfer to real contact-center transcripts has not been piloted, so the literature-grounded risk discussion is not a substitute for a production pilot.The evaluated benchmarks include acted dialogue, scripted TV drama, and unscripted vlogs.Source: §5 · p. 6

Implementation and next steps

  • A cold-start simulation found that 200 hand-labeled utterances recovered full-data oracle weighted F1 within 0.3–0.6 points on average and usually preserved the hybrid’s advantage over both endpoints.The simulation selected the threshold from a small labeled seed sample and evaluated on the disjoint remainder across all three datasets.Source: §5 · p. 6

6 Conclusion

The conclusion argues that a confidence-gated hybrid addresses dataset-dependent differences between a lightweight ensemble and off-the-shelf LLM prompting, while retaining a lower-cost, interpretable routing policy.

6 Conclusion

  • The hybrid Pareto-dominates both pure systems on IEMOCAP, MELD, and CMU-MOSI, across two LLM providers, with accuracy margins statistically confirmed on two datasets.On CMU-MOSI, the smallest test split, accuracy is statistically indistinguishable and the advantage is cost rather than a confirmed accuracy gain.Source: §6 · p. 6
  • The ensemble is not a uniform winner: which system performs better depends on the dataset, motivating a hybrid rather than a single default.Source: §6 · p. 6
  • The router escalates 45.4–58.6% of traffic and remains interpretable enough for CCaaS operations teams to audit.The conclusion characterizes confidence-gated cascading as a practical, low-risk, cost-effective default for conversational-AI emotion recognition.Source: §6 · p. 6

Ethical Statement

The paper uses public benchmark data, reports a limited fairness check, and qualifies the transferability of its results. A hyperparameter search supports the original configuration, while revealing validation overfitting and a data-imbalance mechanism behind CoT’s reversal.

Ethical Statement

  • The study uses only public IEMOCAP, MELD, and CMU-MOSI data, collects no new human-subjects data, and does not redistribute IEMOCAP.IEMOCAP requires a signed release from its maintainers; LLM calls use commercial APIs under standard usage terms.Source: Ethical Statement · p. 6
  • Fairness evidence is limited to IEMOCAP: escalation rates were statistically indistinguishable for female (45.3%) and male (45.6%) speakers, with no significant hybrid weighted-F1 gap.MELD and CMU-MOSI lack usable comparable demographic metadata, so the result is no detected disparity in the one testable case, not general fairness clearance.Source: Ethical Statement · p. 7
  • The ensemble is inspired by Parida et al. (2023), not a reconstruction of that implementation, so production teams should validate against their own baseline.Code and trained artifacts are available to researchers on reasonable request rather than under an open-source license because production training artifacts touch Five9 IP.Source: Ethical Statement · p. 7

A Extended Limitations

  • Provider testing is not a clean replication: MELD’s LLM-wins result did not recur with Llama-3-8B, and larger models such as GPT-4o or Llama-3-70B remain untested.The main experiments use GPT-4o-mini; the paper also notes that its results trail fine-tuned or fusion state-of-the-art systems on IEMOCAP and MELD.Source: A Extended Limitations · p. 8
  • Thresholds τ* are selected retrospectively on fully labeled test splits, so deployment on a new unlabeled dataset requires a cold-start procedure.Appendix E is identified as providing and simulating such a procedure on all three datasets.Source: A Extended Limitations · p. 8

B Hyperparameter Search Validation

  • The original hyperparameters were already close to optimal: tuning improved held-out test weighted F1 by only 0.17–0.78 points, despite internal validation gains of 3–4 points.The validation-to-test gap indicates search overfitting, made more pronounced by using a single holdout per trial to control search cost; the main comparisons and escalation analysis did not change after tuning.Source: B Hyperparameter Search Validation · p. 9, Table 7 · p. 9
  • The tested multi-party attribution hypothesis was unsupported: CoT-versus-few-shot accuracy gaps were statistically indistinguishable across low- and high-speaker-count buckets (0.122 vs. 0.129).The analysis counted distinct speakers in MELD’s preceding five turns; the gap was also non-monotonic across counts of zero through five speakers.Source: B Hyperparameter Search Validation · p. 9, Table 8 · p. 9
  • CoT’s neutral under-prediction interacts with class imbalance: on MELD, true-neutral error rises from 16.9% to 52.5%, while on IEMOCAP it rises from 23.7% to 38.5%.IEMOCAP’s neutral class makes up 24% of the test set versus 48% on MELD, so the same bias costs less weighted F1 there; this is not attributed to MELD’s multi-party structure.Source: B Hyperparameter Search Validation · p. 9, Table 9 · p. 9

D Domain-Shift Risk: Evidence for Transfer to Contact-Center Data

The paper presents literature-based arguments for possible transfer to contact-center data, but its benchmarks are not contact-center datasets and it reports no pilot deployment.

D Domain-Shift Risk: Evidence for Transfer to Contact-Center Data

  • On two evaluated benchmarks, realistic ASR error rates did not hurt IEMOCAP accuracy in Li, Bell, and Lai (2024), but contact-center acoustic conditions remain untested.Their best IEMOCAP ASR transcript at 12.31% word error rate scored 74.66% 4-class accuracy versus 74.32% on ground-truth text; substantial degradation appeared near 40% word error rate.Source: D Domain-Shift Risk: Evidence for… · p. 10
  • The ensemble-to-LLM cascade is a partial structural hedge against domain shift, not a guarantee: shifted inputs may leave ensemble confidence confidently wrong.The paper argues that least-confident ensemble inputs are routed to the component literature suggests is more shift-robust, while explicitly noting calibration under genuine distribution shift is not assured.Source: D Domain-Shift Risk: Evidence for… · p. 10
  • Contact-center language may disrupt feature transfer, and whether escalation actually rises on out-of-domain inputs requires validation in a production pilot.Source: D Domain-Shift Risk: Evidence for… · p. 10
  • Contact-center emotion recognition is an active research area, but evidence that contextual signals help on CusEmo does not establish that this paper’s specific recipe transfers unchanged.The authors identify pilot data from an actual production dataset as future work before making a deployment decision based on these results alone.Source: D Domain-Shift Risk: Evidence for… · p. 10

E Cold-Start Threshold Selection for New, Unlabeled Datasets

With 200 labeled examples, cold-start threshold selection nearly recovers the oracle; further analyses characterize which turns are escalated, the ensemble’s calibration, minority-class routing, and audio’s effect.

E Cold-Start Threshold Selection for New, Unlabeled Datasets

  • 200 labeled seed examples recover the oracle’s weighted F1 to within 0.3–0.6 points on average and preserve hybrid wins in 76–88% of cold starts.At n=200, MELD preserves the “beats both” property in 88% of simulations; its 64% at n=50 reflects greater label needs under class imbalance.Source: E Cold-Start Threshold Selection for… · p. 10, Table 10 · p. 11
  • A stratified sample of at least 200 labels is a starting point; severe imbalance may require more, and a 500-example CMU-MOSI seed leaves a noisy 186-example holdout.The practical recipe recommends periodically re-sweeping τ as QA labels accumulate; operators can select by weighted F1 or a cost-weighted objective.Source: E Cold-Start Threshold Selection for… · p. 11

F Low-Confidence Turn Characterization

  • Emotion or sentiment shifts are the only turn-level property with a consistent escalation-rate direction across all three datasets; other examined properties vary or lack a stable direction.Longer utterances and minority-class membership show different directions across datasets, while speaker-initial position has no stable direction.Source: F Low-Confidence Turn Characterization · p. 11

G Calibration Analysis

  • All three ensemble ECEs fall below 0.05 without explicit calibration, and positive calibration gaps indicate mild under-confidence.ECE is measured on each dataset’s full test split using the ensemble’s own predictions; under-confidence can send already-correct predictions to the LLM rather than retain falsely confident errors.Source: Table 11 · p. 11, G Calibration Analysis · p. 11

H Adaptive Per-Class Thresholds

  • Adding minority-class probability-mass gating raises MELD disgust/fear F1 from 0.000 to 0.167 and macro F1 by 4.1 points, at some cost to weighted F1 and escalation traffic.This option is aimed at operators prioritizing minority-class recall over aggregate weighted F1.Source: H Adaptive Per-Class Thresholds · p. 11

I Audio+Text (Multimodal) Ensemble

  • Audio features widen the ensemble’s IEMOCAP advantage by 3.2 weighted-F1 points and shrink the optimal hybrid escalation rate roughly 5× while matching hybrid accuracy.The added features produce only a marginal positive change on CMU-MOSI and do not reverse the paper’s main comparisons; MELD is not covered because audio is unavailable.Source: I Audio+Text (Multimodal) Ensemble · p. 11

J Cross-Provider Robustness

Testing a second provider finds that dataset rankings are not uniformly provider-independent, but the hybrid’s advantage over pure endpoints replicates across the tested combinations.

J Cross-Provider Robustness

  • The independent-provider evaluation uses Llama-3-8B-Instruct via AWS Bedrock on the complete test splits of all three datasets.The protocol tests whether findings reflect the ERC task or a single LLM provider.Source: J Cross-Provider Robustness · pp. 11–12
  • IEMOCAP favors the ensemble and CMU-MOSI favors Llama-3-8B across strategies, while MELD’s GPT-4o-mini wins do not replicate with Llama-3-8B.On CMU-MOSI the few-shot gap is not significant (p = .140); on MELD, Llama-3-8B loses to the ensemble on every tested strategy.Source: Table 12 · p. 12, J Cross-Provider Robustness · p. 12
  • Across all nine new dataset/strategy combinations, the hybrid matches or beats both pure systems; on IEMOCAP zero-shot it essentially matches the ensemble at +0.0002 weighted F1.That operating point escalates 8.2% of traffic, and the result is consistent with little complementary error when one pure system is close to dominant.Source: Table 13 · p. 12, J Cross-Provider Robustness · p. 12

K Does the Hybrid-Routing Mechanism Generalize to a Stronger Base Classifier?

Replacing the ensemble with a stronger DistilBERT base does not reliably preserve the hybrid’s advantage on every dataset; comparisons to published systems also show that the lightweight approach trails state-of-the-art results.

K Does the Hybrid-Routing Mechanism Generalize to a Stronger Base Classifier?

  • On IEMOCAP, routing over B1 yielded +0.0085 at a 19.5% escalation rate (p = .224), a statistically indistinguishable margin.B1 is a fine-tuned DistilBERT encoder over speaker-tagged context, used as a drop-in base for the unmodified router; its standalone IEMOCAP weighted F1 is 0.618 versus the ensemble’s 0.595.Source: K Does the Hybrid-Routing Mechanism… · pp. 12–13
  • On MELD, the hybrid escalates 59% of traffic and beats the stronger pure LLM significantly (p = .012), largely deferring to that stronger system rather than exploiting complementary errors.At this escalation rate, blended cost approaches pure-LLM cost, weakening the cost-efficiency framing for this pairing.Source: K Does the Hybrid-Routing Mechanism… · p. 13
  • Across three CMU-MOSI seeds, routing improves on B1 but does not reliably beat the pure LLM; fixed-τ escalation varies from 12.8% to 82.1%.B1 recovers 2/30 neutral examples while the LLM recovers 13–26/30, a documented class-level complementarity; deployment with a fine-tuned neural base would need per-model τ recalibration.Source: K Does the Hybrid-Routing Mechanism… · p. 13, Table 15 · p. 13
  • Routing’s benefit depends on a documented base-classifier failure mode that the LLM does not share, and on confidence serving as a usable proxy for that failure.The authors scope the claim to diagnosed dataset- and classifier-specific error patterns, rather than routing architectures in general.Source: K Does the Hybrid-Routing Mechanism… · p. 13

L Comparison to Published State-of-the-Art

  • The audio-augmented IEMOCAP result (.626) trails COSMIC by 2.7 points, while the text-only MELD result (.561) trails Joyful by 5.7 and MiSTER-E by 13.4 points.The MELD comparison is not modality-matched; the gap also remains against largely text-based COSMIC at 9.1 points.Source: Table 16 · p. 13, L Comparison to Published… · p. 14

M Reproducibility

Five-seed checks show small ensemble variation relative to the reported comparison gaps, alongside disclosure of the single-seed headline setup, computing environment, and code-access conditions.

M Reproducibility

  • Code, trained models, and analysis scripts are available to reviewers and researchers upon reasonable request, but the authors do not commit to unconditional public release at submission.The materials include scripts and outputs supporting the reported tables and analyses.Source: M Reproducibility · p. 14
  • Training and analysis ran on a single 16-logical-core, 24 GB RAM Windows machine without a GPU; the ensemble requires no GPU at training or inference.The environment used an AMD Ryzen AI 7 PRO 350.Source: M Reproducibility · p. 14
  • Headline results use one run at seed 42; the authors retrained the text-only ensemble with five seeds per dataset to assess training noise.The ensemble’s stochastic components use the seed argument, while its internal 5-fold split is fixed and unshuffled.Source: M Reproducibility · p. 14
  • Across five training runs, ensemble score spread is small relative to the paper’s main comparison gaps.IEMOCAP has the largest spread (std 0.0067–0.0069), versus 0.0014–0.0025 on MELD and CMU-MOSI.Source: Table 17 · p. 14, M Reproducibility · p. 14

Related papers

Loading 2609.17977v1…