Audio2Tool - Needle-in-a-Haystack: leaderboard
Speech-driven tool calling on Audio2Tool (2026), smart car, smart home and wearable requests spoken by diverse synthetic voices with background noise and given to an end-to-end speech language model, on the needle-in-a-haystack tier (tool requests buried in long irrelevant speech). Exact match of the tool calls and all arguments, in percent; higher is better.
Metric: Exact match (%): share of queries whose predicted tool or ordered tool sequence and all arguments exactly match the reference after deterministic normalization, on the 4,560 Tier 5 needle-in-a-haystack queries (tool calls inside long irrelevant context) of Audio2Tool, with the spoken query (voice-cloned TTS speech mixed with background noise) given to an end-to-end speech language model, zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen2.5-Omni-7B | 40.2 | #1884 |
| 2 | Kimi-Audio | 38.6 | #1735 |
| 3 | Qwen 3 Omni 30B A3B | 36 | #847 |
| 4 | Step-Audio-2-mini | 35.4 | |
| 5 | Audio Flamingo 3 | 28.5 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=audio2tool-needle-in-a-haystack · How It Works · Data refreshed daily, snapshot 2026-11-21.