Datasets:

arcada-labs
/

grocery-bench

language:
  - en
license: mit
pretty_name: Grocery Bench
tags:
  - audio
  - benchmark
  - speech-to-speech
  - voice-ai
  - multi-turn
  - tool-use
  - evaluation
  - state-tracking
  - function-calling
task_categories:
  - automatic-speech-recognition
  - text-generation
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: metadata.jsonl

Grocery Bench

30-turn multi-turn speech-to-speech benchmark for evaluating voice AI models as a grocery ordering assistant.

Part of Audio Arena, a suite of 6 benchmarks spanning 221 turns across different domains. Built by Arcada Labs.

Leaderboard | GitHub | All Benchmarks

Dataset Description

The model acts as a grocery ordering assistant helping a customer build, modify, and finalize an order. The conversation is designed around 15 difficulty enhancements that stress-test item lookup, quantity math, chained corrections, and order reconciliation — culminating in a full order summary the model must compute correctly.

What This Benchmark Tests

Tool use: 5 functions — item lookup, add to cart, remove from cart, modify quantity, order summary
3-item turns: Multiple items added in a single spoken request
Relative-math quantity: "Double the bananas", "add three more"
Conditional addition/removal: "If X costs more than $5, remove it"
Chained corrections: Multiple sequential edits to the same item
Homophone collisions: flower vs flour — ambiguous in speech
Fifteen/fifty audio confusion: Quantities that sound alike over audio
Ambiguous "both": References to multiple items where "both" is under-specified
Revert removal: Undoing a previously removed item
Swap operations: Replace one item with another in a single turn
Retroactive quantity change: Changing a quantity set many turns earlier
Mid-sentence self-correction: Speaker changes their mind partway through
False memory traps: Assertions about items never added
Full order reconciliation: Final order summary requiring correct math across all modifications

Dataset Structure

grocery-bench/
├── audio/                          # TTS-generated audio (1 WAV per turn)
│   ├── turn_000.wav
│   ├── turn_001.wav
│   └── ... (30 files)
├── real_audio/                     # Human-recorded audio
│   ├── person1/
│   │   └── turn_000.wav ... turn_029.wav
│   └── person2/
│       └── turn_000.wav ... turn_029.wav
├── benchmark/
│   ├── turns.json                  # Turn definitions with golden answers
│   ├── hard_turns.json             # Same as turns.json but input_text=null (audio-only)
│   ├── tool_schemas.json           # Tool/function schemas (5 tools)
│   └── knowledge_base.txt          # Grocery store KB (products, policies, delivery)
└── metadata.jsonl                  # HF dataset viewer metadata

Metadata Fields

Field	Description
`file_name`	Path to the audio file
`turn_id`	Turn index (0–29)
`speaker`	`tts`, `person1`, or `person2`
`input_text`	What the user says (text transcript)
`golden_text`	Expected assistant response
`required_function_call`	Tool call the model should make (JSON, nullable)
`function_call_response`	Scripted tool response (JSON, nullable)
`categories`	Evaluation categories for this turn
`subcategory`	Specific sub-skill being tested
`scoring_dimensions`	Which judge dimensions apply

Audio Format

Format: WAV, 16-bit PCM, mono
TTS audio: Generated via text-to-speech
Real audio: Human-recorded by multiple speakers, same transcript content

Usage

With Audio Arena CLI

pip install audio-arena  # or: git clone + uv sync

# Run with a text model
uv run audio-arena run grocery_bench --model claude-sonnet-4-5 --service anthropic

# Run with a speech-to-speech model
uv run audio-arena run grocery_bench --model gpt-realtime --service openai-realtime

# Judge the results
uv run audio-arena judge runs/grocery_bench/<run_dir>

With Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("arcada-labs/grocery-bench")

Evaluation

Models are judged on up to 5 dimensions per turn:

Dimension	Description
`tool_use_correct`	Correct function called with correct arguments
`instruction_following`	User's request was actually completed
`kb_grounding`	Claims are supported by the knowledge base or tool results
`state_tracking`	Consistency with earlier turns (scored on tagged turns only)
`ambiguity_handling`	Correct disambiguation (scored on tagged turns only)

For speech-to-speech models, a 6th turn_taking dimension evaluates audio timing correctness.

See the full methodology for details on two-phase evaluation, penalty absorption, and category-aware scoring.

Part of Audio Arena

Benchmark	Turns	Scenario
Conversation Bench	75	Conference assistant
Appointment Bench	25	Dental office scheduling
Assistant Bench	31	Personal assistant
Event Bench	29	Event planning
Grocery Bench (this dataset)	30	Grocery ordering
Product Bench	31	Laptop comparison shopping

Citation

@misc{audioarena2026,
  title={Audio Arena: Multi-Turn Speech-to-Speech Evaluation Benchmarks},
  author={Arcada Labs},
  year={2026},
  url={https://audioarena.ai}
}