arcada-labs/audio-agent-bench-suite

Audio Agent Bench Suite

A suite of six multi-turn, multi-domain spoken conversational benchmarks for evaluating voice AI and audio agent systems. Each sub-dataset targets a distinct real-world deployment domain, together covering the core capabilities required of production audio agents: instruction following, knowledge-base grounding, tool/function-call accuracy, long-range conversational memory, and state tracking.

Sub-datasets

Dataset	Domain	Turns	HuggingFace
conversation-bench	AI conference assistant (AI Engineer World's Fair)	75	arcada-labs/conversation-bench
product-bench	Laptop sales assistant (TechMart Electronics)	31	arcada-labs/product-bench
grocery-bench	Grocery ordering assistant (Harvest & Hearth Market)	30	arcada-labs/grocery-bench
appointment-bench	Dental office scheduling assistant (Bayshore Family Dental)	25	arcada-labs/appointment-bench
event-bench	Event planning assistant (Evergreen Events)	29	arcada-labs/event-bench
assistant-bench	Personal assistant with flights, hotels, calendar & email (Atlas)	31	arcada-labs/assistant-bench

Overview

Each sub-dataset consists of a sequence of conversational turns designed to be evaluated as a complete multi-turn dialogue. Every turn contains:

• A human voice recording (WAV) of the user's input
• A transcript of the spoken input
• A golden reference response that an ideal agent should produce
• Function/tool call specifications for turns requiring tool use (null otherwise)
• Category labels and scoring dimensions for structured evaluation

The benchmark is English-only and is intended for evaluation purposes only, not for training.

Schema

All six sub-datasets share the following schema:

Field	Type	Description
turn_id	int	Sequential turn index (0-based)
input_text	string	Transcript of the user's spoken input
golden_text	string	Reference golden response
required_function_call	list or null	Function/tool calls required for this turn
function_call_response	list or null	Expected responses from function calls
categories	list	Turn-level category labels
subcategory	string or null	Optional finer-grained category
scoring_dimensions	list	Evaluation dimensions applicable to this turn
audio_file	string	Relative path to the WAV audio file

Categories

Category	Description
basic_qa	Factual question answering from the knowledge base
tool_use	Requires calling one or more functions/tools
long_range_memory	Requires recall of information from earlier in the conversation
state_tracking	Involves cross-entity references, chained corrections, or multi-step state changes
ambiguity_handling	Input is ambiguous and requires disambiguation or appropriate clarification

Scoring Dimensions

Each turn is evaluated across up to 5 dimensions. Core dimensions (tool_use_correct, instruction_following, kb_grounding) are scored on every turn; state_tracking and ambiguity_handling are scored only on turns tagged with the relevant categories.

Dimension	Scored on	Description
tool_use_correct	All turns	Whether required tool/function calls were made correctly. Missed calls land here unless a more specific dimension applies.
instruction_following	All turns	Whether the assistant's words and actions followed the user's instructions in a non-tool sense. Strictly separated from tool_use_correct.
kb_grounding	All turns	Whether the response is grounded in the knowledge base provided in the system prompt.
state_tracking	Tagged turns	Whether the agent correctly tracked conversational state across turns (e.g. cross-entity references, chained corrections, long-range memory).
ambiguity_handling	Tagged turns	Whether the agent correctly resolved ambiguous input — neither over-clarifying nor hallucinating a resolution. Over-clarification penalties land here.

Loading the Data

Each sub-dataset can be loaded individually using the HuggingFace datasets library:

from datasets import load_dataset

# Load a specific sub-dataset
ds = load_dataset("arcada-labs/conversation-bench")

# Load all six
domains = [
    "conversation-bench",
    "product-bench",
    "grocery-bench",
    "appointment-bench",
    "event-bench",
    "assistant-bench",
]
datasets = {name: load_dataset(f"arcada-labs/{name}") for name in domains}

Each sub-dataset has a default configuration with a train split containing all benchmark turns.

Audio Files

Audio files are recordings of 2 human English-speaking voice actors reading the scripted user inputs. They are stored in the audio/ directory of each sub-dataset repo and referenced by relative path in the audio_file field (e.g. audio/turn_000.wav).

Evaluation

The benchmark is designed for end-to-end evaluation of audio agent systems: the audio file is the input, and the system's response is compared against golden_text along the applicable scoring_dimensions. For tool-use turns, required_function_call and function_call_response define the expected tool interaction.

Responsible AI

This dataset was collected and annotated with the following considerations:

• Fictional identities: All names used in scripts (e.g. "Jennifer Smith") are fictional. No real personally identifiable information appears in the text modality.
• Speaker consent: Audio recordings feature 2 human English-speaking voice actors who provided consent for inclusion of their recordings.
• No sensitive data: No financial, health, or other sensitive personal data is present.
• Evaluation only: The dataset is not intended for model training.
• Known limitations: English-only; limited to predefined domains; audio features only 2 human speakers, which may not reflect the acoustic diversity of real-world users (accents, speaking styles, noise conditions); small scale (25–75 turns per sub-dataset).

License

Creative Commons Attribution 4.0

Citation

If you use the Audio Agent Bench Suite in your research, please cite:

@dataset{arcada_labs_audio_agent_bench_suite,
  author    = {Arcada Labs},
  title     = {AudioAgentBench: Evaluating Multi-Turn Voice Agents on Real-World Tasks},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/arcada-labs/audio-agent-bench-suite}
}