Whisper-v3-large + GPT-4o vs GPT-4o-Audio

Compare their published VoiceBench columns before drawing conclusions about a deployment.

Third-party Benchmark · VoiceBench · Captured 2026-09-27

These are published model results. VoiceAgent.best ran no model calls or voice-agent tests. They do not measure a platform, phone call, production latency, voice quality or handoff reliability.

Leaderboard · Repository · Paper · Dataset card · Apache-2.0 repository terms

Citation: Chen, Yue, Zhang, Gao, Tan and Li (2024). VoiceBench: Benchmarking LLM-Based Voice Assistants. arXiv:2410.17196.

Shared published evidence

Both rows have 9 populated dimensions in the same captured source table. Their metric columns and display scales match. Exact evaluation configurations, API checkpoints and run dates remain unknown, so this is a comparison of published rows rather than a controlled experiment.

Two model rows, same VoiceBench capture and published metric scales.
Third-party Benchmark · VoiceBench · 2026-09-27 · Exact checkpoint and protocol version unknown.
Model	Architecture	Weights	AlpacaEvalsource 5-point scale	CommonEvalsource 5-point scale	WildVoicesource 5-point scale	SD-QAsource score /100	MMSUsource score /100	OBQAsource score /100	BBHsource score /100	IFEvalsource score /100	AdvBenchsource score /100	VoiceBench OverallSource reported /100	Provenance
GPT-4o-Audio	OmniCommunity label	closedSource classification	4.78	4.49	4.58	75.50	80.25	89.23	84.10	76.02	98.65	86.75VoiceBench Overall	VoiceBenchThird-party Benchmark 2026-09-27
Whisper-v3-large + GPT-4o	CascadedCommunity label	closedSource classification	4.80	4.47	4.62	75.77	81.69	92.97	87.20	76.51	98.27	87.80VoiceBench Overall	VoiceBenchThird-party Benchmark 2026-09-27

Qwen3-Omni rows import official results: the leaderboard divides the three open-ended QA values by 20 and rounds to two decimals; Overall remains as officially reported. We preserve these published values and create no universal score. Run settings and checkpoint equivalence are unverified.

Reasoning, instruction following and safety

Reasoning

Whisper-v3-large + GPT-4o: 87.20
GPT-4o-Audio: 84.10

BBH · source score /100 · Third-party Benchmark · 2026-09-27

Instruction following

Whisper-v3-large + GPT-4o: 76.51
GPT-4o-Audio: 76.02

IFEval · source score /100 · Third-party Benchmark · 2026-09-27

Safety

Whisper-v3-large + GPT-4o: 98.27
GPT-4o-Audio: 98.65

AdvBench · source score /100 · Third-party Benchmark · 2026-09-27

How to use this comparison

Use the dimensions to identify questions for a real pilot: which instructions fail, which reasoning tasks matter and which unsafe requests should be escalated. The table has no measurements of call interruption, network behavior, cost or customer outcomes.

Differences cannot be attributed to architecture alone. Other model, prompt and evaluation choices are uncontrolled in this public snapshot. Supporting one of these model families does not transfer these scores to a platform.