A Reliability Assessment of LALM Audio Judges for Full-Duplex Voice Agents¶
π Published (v1): 2026-07-08 23:24 UTC Β· Source: Arxiv Β· link
Why this paper was selected
Empirical reliability of audio LLM judges for full-duplex voice agents β evaluation methodology for conversational agents
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
π¬ Ask ChatGPTβ¦ Ask Claude
TL;DR¶
This paper empirically validates Gemini audio-language models (LALMs) as judges of full-duplex voice-agent conversations, scoring raw stereo waveforms on 8 production dimensions against a 3-rater human reference standard. On 5β6 of 8 dimensions, Gemini 2.5 Flash performs comparably to a fourth human rater, at roughly two orders of magnitude lower cost. The study is the first to validate a LALM judge over multi-rater, full-duplex conversational audio rather than isolated TTS utterances.
Problem¶
Prior LALM-as-judge work benchmarks models on isolated text-to-speech utterances with single-rater or system-level MOS correlations. No study had validated a LALM judge against a calibrated multi-rater human standard on raw stereo full-duplex agentβclient conversations, where turn-taking, channel asymmetry, and conversational disfluencies all affect rating distributions. The study fills this gap for production-grade evaluation pipelines.
Method¶
Corpus: 209 stereo sessions from a production customer-support voice agent β 152 natural full-duplex conversations across 13 accent-and-condition strata (6 clean Γ 6 accents, plus 7 codec-degraded strata), plus 57 adversarial defect-injected clips (6 DSP defects: hard clipping, dead air, additive noise at β2 to β8 dB SNR, truncation, phoneme overdub, 8 kHz resample).
Judges: Two production LALM judges (AgentSpeechFidelity and ConversationalAudioQuality, both Gemini 2.5 Flash via Vertex AI) receive raw stereo WAV bytes with no transcription; each returns structured JSON scores on 8 Likert-scale (1β5) dimensions.
Human baseline: 3 Salesforce-contracted annotators scored all 209 sessions Γ 8 dimensions with anchored rubrics and a pre-batch calibration meeting (5,016 observations total).
Agreement statistics: (i) pairwise Spearman \(\rho\) (1,000-iteration bootstrap 95% CIs) for rank agreement; (ii) proportion of sessions within 1 point of the 3-rater human mean; (iii) ordinal Krippendorff \(\alpha\) for chance-corrected reliability; (iv) Newcombe-Wilson CIs on LALM-minus-human recall difference for defect detection.
Cross-model replication: All 209 sessions re-scored with Gemini 3.5 Flash and Gemini 3.1 Pro Preview, same prompts, same judges.
Key Contributions¶
- First multi-rater validation of a LALM judge on raw full-duplex stereo agent-client audio (vs. prior work on isolated utterances).
- Empirical evidence that Gemini 2.5 Flash achieves near-human rank agreement (\(\rho\) gap \(\leq 0.07\)) on 5/8 dimensions and \(\geq\)60% within-1-point simple agreement on 6/8 dimensions.
- Identification of a "reliability paradox": Krippendorff \(\alpha \approx 0\) even at >90% within-1 agreement on ceiling-saturated dimensions, motivating per-regime reporting.
- Adversarial defect analysis showing LALM and human coverage are complementary, not nested (e.g., LALM catches clipping as prosody degradation; humans catch it as
audio_claritydegradation). - Cross-model finding that rank-ordering transfers across Gemini versions, but absolute calibration does not β a model swap requires re-validating simple agreement, not just rank correlation.
- Directional cost estimate: ~\(2\) orders of magnitude cost reduction vs. three-rater human coverage of a 100-session weekly batch.
Results¶
- Rank agreement: LALM-human \(\rho\) departs from human-human \(\rho\) by \(\leq 0.07\) on 5/8 dimensions; 95% bootstrap CIs overlap on 7/8 dimensions for Gemini 2.5 Flash.
- Simple agreement: \(\geq\)60% within-1-point of human mean on 6/8 dimensions; three dimensions reach \(\geq\)89% (
accent_dialect_handling92.3%,speech_clarity90.9%,audio_clarity89.0%). - Failing dimensions:
speaking_rate_adaptation(48.8%) andoverall_fidelity(43.5%) fall below 50%; humans themselves show low rank agreement there (Ο = 0.17 and 0.11). - Adversarial sensitivity: LALM is significantly more sensitive than humans in 4/48 (defect Γ dimension) cells; humans are significantly more sensitive in 3/48 cells; 41/48 show no significant difference (underpowered at \(n \leq 8\)).
- Critical blind spot:
clip Γ audio_clarityβ humans detect clipping 8/8; LALM 0/8. Robust under paired McNemar test. - Gemini 3.5 Flash: Improves simple agreement to 8/8 dimensions (\(\geq\)60%), rank-correlation gap within 0.07 on 4/8.
- Gemini 3.1 Pro Preview: Rank correlation comparable (5/8 within 0.07), but calibration offset causes
audio_claritymean score to be β1.52 points below human mean; within-1 agreement drops to 34.9% onaudio_clarityand 37.3% onspeaking_rate_adaptation. - Natural conversations alone (n=152): Headline counts unchanged;
audio_clarityhuman-human \(\rho\) collapses from 0.47 to 0.07 once adversarial clips are excluded, revealing pooled correlation was measuring shared defect sensitivity, not natural-speech discrimination.
Limitations¶
- Primary evidence base is a single LALM (Gemini 2.5 Flash) on a single production system; generalizability to other voice agents or LALM providers is untested.
- Adversarial arm is underpowered: \(n \leq 8\) per (defect, dimension) cell β 41/48 non-significant cells are absence of power, not demonstrated equivalence.
- Human raters are a single three-person panel from one annotation vendor; inter-rater variance is high (rater means differ by >1 Likert point on
speech_clarityandaudio_clarity), limiting the robustness of the human ground truth. - Cost estimates are directional internal figures, not reproducible unit-cost models.
overall_fidelityandspeaking_rate_adaptationshow poor reliability for both humans and LALM; the paper attributes this to rubric construct invalidity but does not fix it.- The study covers only the Gemini model family; cross-provider generalization is unexamined.
- Paper text is truncated before Section 7 is complete, so additional stated limitations may exist.
Relevance to Agentic AI / LLM Agents¶
Automated evaluation of voice agents is a direct bottleneck for scaling agentic AI quality assurance: human rating is the current ceiling on how frequently production agent conversations can be audited. This work provides a methodology and empirical threshold (within-1 simple agreement \(\geq\)60%, \(\rho\) gap \(\leq 0.07\)) for deciding when an LLM-as-judge can substitute for human raters in production agent pipelines, directly applicable to any agentic system producing auditable outputs. The cross-model calibration finding β that rank-ordering transfers but absolute scoring does not across model versions β is a practical warning for teams maintaining automated judge pipelines for agents as underlying models are upgraded. The complementary-coverage pattern (LALM and humans catch different defects) argues for hybrid human-LALM evaluation architectures rather than full replacement, a design principle relevant to any agent monitoring system.