Datasets:
scenario stringclasses 1
value | row_idx int32 1 100 | call_id int32 408 773 | user_id int32 2.24k 3.12k | provider stringclasses 5
values | scenario_description stringclasses 1
value | duration_sec float32 121 774 | merged_audio audioduration (s) 121 774 | user_audio audioduration (s) 121 774 | agent_audio audioduration (s) 121 774 | sampling_rate int32 24k 48k | merged_channels int32 2 2 | tool_calls_json stringlengths 14.4k 385k | num_tool_calls int32 3 158 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Scenario 1 | 1 | 551 | 2,242 | human | Book a Flight for Two People | 439.182007 | 44,100 | 2 | "[{\"ts\":\"22:51:00\",\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BO(...TRUNCATED) | 158 | |||
Scenario 1 | 2 | 515 | 2,242 | grok | Book a Flight for Two People | 486.658997 | 48,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BOM\",\"date\":\"2026-(...TRUNCATED) | 4 | |||
Scenario 1 | 3 | 537 | 2,242 | inworld | Book a Flight for Two People | 435.130005 | 48,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BOM\",\"date\":\"2026-(...TRUNCATED) | 7 | |||
Scenario 1 | 4 | 540 | 2,242 | gemini | Book a Flight for Two People | 417.881989 | 24,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"date\":\"2026-07-21\",\"origin\":\"DEL\",\"destination\":(...TRUNCATED) | 4 | |||
Scenario 1 | 5 | 544 | 2,242 | openai | Book a Flight for Two People | 399.799988 | 48,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BOM\",\"date\":\"2026-(...TRUNCATED) | 5 | |||
Scenario 1 | 6 | 527 | 2,244 | human | Book a Flight for Two People | 476.890991 | 44,100 | 2 | "[{\"ts\":\"22:51:00\",\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BO(...TRUNCATED) | 123 | |||
Scenario 1 | 7 | 514 | 2,244 | gemini | Book a Flight for Two People | 216.462006 | 24,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"date\":\"2026-07-21\",\"origin\":\"DEL\",\"destination\":(...TRUNCATED) | 7 | |||
Scenario 1 | 8 | 519 | 2,244 | openai | Book a Flight for Two People | 362.649994 | 48,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BOM\",\"date\":\"2026-(...TRUNCATED) | 4 | |||
Scenario 1 | 9 | 578 | 2,244 | grok | Book a Flight for Two People | 512.672974 | 48,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BOM\",\"date\":\"2026-(...TRUNCATED) | 4 | |||
Scenario 1 | 10 | 565 | 2,244 | inworld | Book a Flight for Two People | 430.309998 | 48,000 | 2 | "[{\"name\":\"search_flights\",\"args\":{\"origin\":\"DEL\",\"destination\":\"BOM\",\"date\":\"2026-(...TRUNCATED) | 4 |
Voice Arena Goal Dataset (Indian English)
200 task-oriented voice calls between human callers and five agents (Gemini, Grok, Inworld, OpenAI and a human agent), with pairwise human judgments of task capability and humanness.
Paper (arXiv:2609.29431) · Citation
VA was introduced in Calibrating LLM Judges for Human and AI Conversations (Züfle et al., 2026). It tests whether an automatic judge can tell which of two spoken conversations was more successful. Each call is a human caller trying to complete an airline task. The same caller makes the same request to all five agents, and human raters compare the calls pair by pair.
| Calls | 200 (100 flight booking, 100 reschedule), 17.75 hours |
| Callers | 20, each completing both scenarios with all five agents |
| Per call | Merged two-channel audio, caller audio, agent audio, tool-call log |
| Judgments | 1,606 pairwise judgments over 397 call pairs, released as per-pair vote counts |
| Labels | Majority winner for task capability and humanness |
| Language | English (Indian English callers) |
Not included: transcripts (generate your own with ASR), model versions, rater-level judgments.
Quick start
The audio is about 3.96 GB. Use streaming=True if you only need part of it.
pip install datasets torchcodec huggingface_hub pandas pyarrow
from datasets import load_dataset
calls = load_dataset("VoiceArena/Goal-Dataset_en_in") # both splits
votes = load_dataset(
"csv",
data_files="hf://datasets/VoiceArena/Goal-Dataset_en_in/pairwise_majority_votes.csv",
split="train",
)
Join judgments to calls on call_id. Pairs with a clear human preference (the evaluation set used in the paper) are those where *_majority_model != "no_majority": 260 pairs for task capability, 299 for humanness.
Structure
| Path | Split | Rows | Content |
|---|---|---|---|
data/flight_booking-*.parquet |
flight_booking |
100 | Book a Flight for Two People |
data/reschedule-*.parquet |
reschedule |
100 | Reschedule One Passenger |
pairwise_majority_votes.csv |
n/a (load as CSV) | 397 | Pairwise human judgments |
The splits are the two scenarios, not a train/test partition. VA is meant for evaluation.
Call fields
| Field | Description |
|---|---|
scenario, scenario_description |
Scenario 1 / Book a Flight for Two People or Scenario 2 / Reschedule One Passenger. |
row_idx |
Row number within the split, 1 to 100. |
call_id |
Unique call ID across both splits. Joins to call_id_a / call_id_b in the judgments file. |
user_id |
Caller ID. Same callers in both splits. Joins to caller_user_id. |
provider |
gemini, grok, inworld, openai or human. |
duration_sec |
Length of the merged recording. |
merged_audio |
Full call, two channels. |
user_audio, agent_audio |
Caller and agent channels separately. |
sampling_rate |
24,000, 44,100 or 48,000 Hz. Resample before batch processing. |
merged_channels |
Always 2. |
tool_calls_json, num_tool_calls |
JSON array of the agent's backend tool calls, and its length. |
Tool-call logs
All five agents used the same simulated airline tools, so logs show whether the agent actually completed the task.
| Scenario | Tools |
|---|---|
| Flight booking | search_flights, list_available_seats, book_flight, process_payment |
| Reschedule | find_booking_by_pnr, search_flights, list_available_seats, reschedule_passenger, collect_reschedule_payment |
| Calls | Keys per entry | Time field |
|---|---|---|
| AI agents | name, args, response, called_at, latency_ms |
called_at: seconds from call start |
| Human agent | ts, name, args, response, status |
ts: wall-clock HH:MM:SS, no date |
Human-agent logs can include actions from other calls, hours apart (up to 158 entries in flight booking, 105 in reschedule). Restrict them to the call's time window before scoring task success.
Pairwise judgments file
One row per judged pair. a and b are ordered by ascending call_id, not by what raters heard first.
| Column | Description |
|---|---|
scenario, caller_user_id |
Scenario and the caller who made both calls. |
call_id_a, model_a, call_id_b, model_b |
The two calls and their agents. |
n_raters |
Raters on the pair (1 to 6). |
task_votes_{a,b,tie} |
Task-capability votes. |
task_majority_call_id, task_majority_model |
Majority winner, or empty / no_majority. |
humanness_votes_{a,b,tie} |
Humanness votes. |
humanness_majority_call_id, humanness_majority_model |
Majority winner, or empty / no_majority. |
Majority rule. A call wins only with more than half of all votes, ties included (3 of 5 wins; 3 of 6 and 2 of 4 do not). Otherwise no_majority.
Agent labels. gemini, grok, inworld, openai, human in the call data appear as Gemini, Grok, Inworld, OpenAI, Human anchor in the judgments file.
Unjudged pairs. 3 of the 400 possible pairs have no judgments: calls 463 vs 496 (caller 2252, flight booking), 680 vs 686 and 680 vs 699 (caller 2246, reschedule).
Statistics
| Flight booking | Reschedule | |
|---|---|---|
| Calls | 100 | 100 |
| Mean duration | 354.3 s | 284.9 s |
| Judged pairs | 199 | 198 |
| Pairs with task majority | 132 | 128 |
| Pairs with humanness majority | 152 | 147 |
Mean call duration by agent (seconds, flight booking / reschedule): Gemini 259.3 / 234.1, OpenAI 310.7 / 266.9, Human agent 342.0 / 292.4, Inworld 421.1 / 309.7, Grok 438.2 / 321.1.
Raters per pair: 1 to 6, mean 4.05. Tie rate in individual judgments: 19.1% (task), 13.6% (humanness).
Majority wins by agent (each agent is in 158 to 160 pairs):
| Agent | Task | Humanness |
|---|---|---|
| Human agent | 124 | 144 |
| Gemini | 89 | 101 |
| OpenAI | 21 | 25 |
| Inworld | 13 | 16 |
| Grok | 13 | 13 |
These are label counts, not a leaderboard. They ignore ties and come from one domain, one locale and 20 callers.
Data collection
Both scenarios use a simulated airline, DEL to BOM on 21 July 2026. In the booking scenario the agent captures passenger count, route, date and seat preference, then books and confirms. In the reschedule scenario the agent finds an existing booking, moves one passenger to a different flight the same day, and confirms.
- Callers. 20 human callers. Each completed both scenarios (reschedule after booking) and called all five agents once per scenario, in randomised order, in one session per scenario.
- Agents. Gemini (Live API), Grok (Voice Agent API), OpenAI (Realtime API), Inworld (Realtime API), and a human agent using the same tools. The four AI agents ran the same system prompt, tools and backend (three fare tiers, economy cabin, prices in rupees). AI calls were recorded 17 July to 1 August 2026.
- Procedure. The caller gets a goal brief (not a script) in the Voice Arena app, calls the agent from inside the app, and supplies details on request (names, dates of birth, phone number, seat and fare choice). The agent uses the airline tools. Both sides are recorded.
Human annotation
Raters heard two calls from the same task instance (same caller, same request, two agents), labelled A and B, audio only, without being told the agent. For each pair they answered:
| Criterion | Question | Answers |
|---|---|---|
| Task capability | Did the agent successfully complete the task? | A, B or tie |
| Humanness | How human-like did the agent sound? | A, B or tie |
Judgments came from 83 raters. Leave-one-rater-out human agreement on the majority-labelled pairs is 84.2% (task) and 86.7% (humanness).
Cleaning (done before release): removed exact duplicates, raters flagged by quality checks (with all their judgments), judgments involving five callers flagged at recording time (their calls are not released), judgments missing a humanness answer, and 15 judgments from a late batch with time on task of 200% or more of the recording length.
Baselines from the paper
Pairwise accuracy (%) on pairs with a majority label, best prompt per model (Table 3). Text results are the mean over Parakeet and Whisper Large V3 transcripts.
| Judge | Input | Task | Humanness |
|---|---|---|---|
| Duration baseline (shorter call wins) | n/a | 65.0 | 60.4 |
| Human ceiling | n/a | 84.2 | 86.7 |
| Phi-4-Multimodal | audio | 49.5 | 67.2† |
| Qwen2.5-Omni-7B | text | 64.7† | 60.0 |
| Qwen3-Omni-30B | audio | 38.4 | 37.9 |
† significantly above chance (95% CI). See the paper for all judges. No judge reaches human agreement, and only one beats the duration baseline on humanness.
Notes
- Scope. One domain (airline), two scenarios, one route and date, one locale, 20 callers. Use VA to evaluate judges, not to estimate how often an agent succeeds in production.
- Duration is a strong signal. Compare judges against the duration baseline, not only against chance.
- Few raters on some pairs. 46 of 397 pairs have one or two raters.
- Blinding. Comparisons between two AI agents are blind. When the human agent is in the pair, the voice may reveal it.
- Call order. Rater listening order and each caller's call order are not released.
Personal and sensitive information
The recordings contain the voices of the callers and the human agent. tool_calls_json includes passenger names, dates of birth and a contact phone number given during the calls. Do not try to identify participants or use the recordings for voice cloning or speaker identification. To request removal of a recording, open a discussion in the Community tab.
Intended and out-of-scope uses
Intended uses
- Evaluating judges of spoken conversational success, pairwise or pointwise.
- Testing whether calibration transfers from human-human to human-AI conversations.
- Research on voice agents: turn-taking, latency, error recovery, tool use.
Out-of-scope uses
- Ranking or marketing commercial voice products from this dataset alone.
- Speaker identification or voice cloning.
- Training systems to imitate the callers or the human agent.
Citation
@misc{züfle2026calibratingllmjudgeshuman,
title={Calibrating LLM Judges for Human and AI Conversations},
author={Maike Züfle and Patrícia Schmidtová and Vilém Zouhar and Shree Harsha Bokkahalli Satish and Erica Cooper and Nicholas Sanders and Shobhit Banga and Vaibhav Nalawade and Manmeet Kaur and Jan Niehues and Markus Müller and Ondřej Klejch},
year={2026},
eprint={2609.29431},
archivePrefix={arXiv},
primaryClass={cs.HC},
url={https://arxiv.org/abs/2609.29431},
}
Supported by JSALT 2026 at Johns Hopkins University. Curated by Voice Arena. Questions or corrections: open a discussion in the Community tab.
- Downloads last month
- 558