Teaming-UBA: Detecting teaming from the teammate's voice

Laboratorio de Inteligencia Artificial Aplicada (LIAA)
Instituto de Ciencias de la Computación (ICC), UBA / CONICET
Buenos Aires, Argentina

Dataset for a study on human–agent teaming through spoken interaction.

Subjects played a collaborative game that required speaking with a virtual agent, and each subject interacted with two agents that provided the same task-relevant information but differed in behaviour: a neutral, tool-like agent (N agent) and a team-building agent (T agent) that used simple social strategies such as empathy, politeness and positivity. The study looks at how the agent's behaviour and the subject's personality traits relate to the perception of the agent as a teammate and to task performance, and whether the agent's behaviour can be detected from the subject's speech.

Protocol

The game, called What-Where, has two roles: the subject is the follower and the agent is the describer. See the paper, currently in the process of publication, for details on how the agent was operated. Each trial has two stages:

• Stage 1 (target identification / "what") — the subject asks questions to identify a target object among 30 similar objects on a board.
• Stage 2 (target localization / "where") — the subject places the identified object at a specific position on a new board, known only to the agent, by asking about its position relative to other objects or the screen.

Before playing with each agent, subjects were shown an introductory text intended to bias their expectations of that agent, which we refer to as the priming: the name of the agent (Iris or Ambar), a rating on three dimensions — ability, communication and cooperation — and comments from presumed previous users. Both agents delivered the same task-relevant information, but the T agent additionally used positive-politeness strategies (greetings, expressions of enthusiasm, encouraging language after a poor outcome, diminutives).

The experiment was run in two versions, which differ only in this priming:

• active-priming, the main version, in which the priming favoured the T agent: the rating on ability was the same for both agents, whereas the ratings on communication and cooperation were lower for the N agent, and the comments additionally described the T agent as more involved, sociable, pleasant and helpful.
• neutral-priming, a second version run with a smaller number of subjects, in which the priming for the T agent was equivalent to that of the N agent. It allows the contribution of the priming to be disentangled from that of the agents' behavioural differences.

Each subject played:

• 4 training trials (always with a neutral agent, not analyzed and not included in the released audio).
• 4 trials with one agent (N or T), then 4 trials with the other. Each trial had one of three difficulty levels: easy, intermediate, difficult.

Surveys

Subjects answered the following questionnaires: the first four on a separate day, before the session in which the game was played, and the AATL ones during that session, after each trial and at the end of each block.

• Sociodemographic questionnaire — age, gender, education, language, accent region, hearing and color vision.
• Big Five Inventory (BFI) — 44 items, 1–5 Likert scale, grouped into 5 factors: openness, conscientiousness, extraversion, agreeableness, neuroticism.
• Interpersonal Orientation Scale (IOS) — 26 items, 1–5 Likert scale, grouped into 4 factors: emotional support, attention, positive stimulation, social comparison.
• Scale of Trust in Automated Systems (STAS) — 12 items, 1–7 Likert scale, single trust-in-automation score.
• Autonomous Agent Teammate-Likeness (AATL) — administered in two forms: a 53-item final questionnaire answered twice per subject, once per agent at the end of each block (7 factors: autonomy, altruism, cooperative, relationship-building, communication, synchronized, ability), and a shorter 4-factor intermediate questionnaire answered after each trial (altruism, communication, relationship-building, ability).

Download

Everything is stored uncompressed, so the tables can be downloaded without the audio.

Tables only: the surveys, the game logs and the alignments.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mgauder/Teaming-UBA",
    repo_type="dataset",
    allow_patterns="dataset/*/dataframes/*.csv",
    local_dir="Teaming-UBA",
)

Everything: the same tables plus the WAV recordings.

snapshot_download(
    repo_id="mgauder/Teaming-UBA",
    repo_type="dataset",
    local_dir="Teaming-UBA",
)

Layout

dataset/
├── active-priming/
│   ├── audios/
│   │   ├── session_289/
│   │   │   ├── block_1-trial_1-stage_1.wav
│   │   │   ├── ...
│   │   │   ├── block_2-trial_4-stage_2.wav
│   │   │   ├── neutral_reading.wav
│   │   │   └── neutral_free_speech.wav
│   │   ├── session_326/
│   │   └── ...                                  (99 sessions)
│   └── dataframes/
│       ├── metadata.csv
│       ├── personal_information_answers.csv
│       ├── Big_Five_Inventory_survey_answers.csv
│       ├── Interpersonal_Orientation_Scale_survey_answers.csv
│       ├── Scale_of_Trust_in_Automated_Systems_survey_answers.csv
│       ├── final_survey_answers.csv
│       ├── intermediate_survey_answers.csv
│       ├── silero_vad_aligns.csv
│       └── neutral_silero_vad_aligns.csv
└── neutral-priming/                             (same structure, 19 sessions)
    ├── audios/
    └── dataframes/

If you download the audio, keep this layout: the dataset_path column of silero_vad_aligns.csv and neutral_silero_vad_aligns.csv is relative to dataset/, so the alignments resolve to the audio files only when both sit under the same root.

The dataset

The dataset holds the two versions of the experiment described above, one per directory:

• active-priming/ — 99 sessions released (61 female, 37 male, 1 other; mean age 32.1, SD 10.2).
• neutral-priming/ — 19 sessions released; 18 of them answered the questionnaires administered on the previous day (14 female, 3 male, 1 other; mean age 36.1, SD 8.9).

Both versions share the same structure. Every session collected is released.

audios/

audios/session_<session_id>/ contains, per subject:

• block_<1-2>-trial_<1-4>-stage_<1-2>.wav — one recording per stage of each of the 8 trials (4 with each agent). block_<1-2> is the chronological order in which the subject played that agent (1: first, 2: second); use metadata.csv (block_order, agent_condition) to know which agent it corresponds to for a given subject.
• neutral_reading.wav — reading of a phonetically balanced text.
• neutral_free_speech.wav — description of a picture (at least 30 seconds).

Recordings are WAV PCM 16-bit, mono or stereo, at 16-48 kHz: sessions were remote and each subject used their own microphone and headset. Every session has its 16 stage recordings; the two neutral recordings are present for all sessions (except session number 357).

dataframes/

All files are UTF-8 CSV with a header row, comma-separated, and no index column.

import pandas as pd

metadata = pd.read_csv("dataset/active-priming/dataframes/metadata.csv")
subjects = pd.read_csv(
    "dataset/active-priming/dataframes/personal_information_answers.csv",
    parse_dates=["dob"],
)

metadata.csv

One row per trial.

Column	Type	Description
session_id	Integer	Subject/session identification. Each subject solves one session.
block_order	Integer (1-2)	Whether this trial belongs to the subject's first or second agent block.
block_type	Char (A/B)	Which of the two mirrored board sets was used in this block.
trial_order	Integer (1-4)	Order of the trial within its block.
trial_difficulty	String (easy / intermediate / difficult)	Difficulty level of the trial.
agent_condition	String (N agent / T agent)	Agent type for this trial.
agent_order	String (N-FIRST / T-FIRST)	Whether this subject played the N agent or the T agent first.
session_block_id	Integer	Identifier of the block this trial belongs to (one per subject × agent block; matches `session_block_id` in `final_survey_answers.csv`).
session_trial_id	Integer	Identifier of this trial; one per row (matches `session_trial_id` in `intermediate_survey_answers.csv`).
first_step_attempts	Integer	Number of attempts in Stage 1 before the subject selected the correct target object.
overlap_score	Float	Stage 2 score: overlap score between the position selected by the subject and the actual target position.
distance_score	Float	Stage 2 score: distance between the position selected by the subject and the actual target position.
string_score	String (bad / fair / good / excellent)	Stage 2 categorical performance label shown to the subject, based on `overlap_score`.
what_discontinuous_task, where_discontinuous_task	Bool	Set if Stage 1 / Stage 2 of this trial was interrupted (e.g. by a connection loss) and resumed.
agent_what_answers_count, agent_where_answers_count	Float	Number of agent responses during Stage 1 / Stage 2 of the trial.

personal_information_answers.csv

One row per subject.

Column	Type	Description
session_id	Integer	Subject/session identification.
dob	Date	Date of birth.
age	Float	Age at the time of the session.
gender	String (male / female / other)	Self-reported gender.
accent	String	Self-reported accent (free text).
first_language	String	Self-reported first language (free text).
born, address	String	Place of birth and current residence (free text).
address_time	String	Time living at the current address, free text as typed by the subject: mostly a number of years, but some answers are written out in Spanish (e.g. 6 meses, Toda la vida).
education	String	Self-reported highest education level, normalized to: secondary_incomplete, secondary_complete, tertiary_incomplete, tertiary_complete, university_incomplete, university_complete, masters_incomplete, masters_complete, doctorate_incomplete, doctorate_complete, unknown.
hearing	Bool	Self-reported hearing impairment.
color_blindness	Bool	Self-reported color blindness.

Big_Five_Inventory_survey_answers.csv, Interpersonal_Orientation_Scale_survey_answers.csv, Scale_of_Trust_in_Automated_Systems_survey_answers.csv

One row per item answered. Same schema for the three files.

Column	Type	Description
session_id	Integer	Subject/session identification.
name	String	Name of the survey.
text, text_english	String	Item text, in Spanish (as shown to the subject) and English.
response	Integer	Subject's answer: 1-5 Likert scale for BFI and IOS; 1-7 for STAS.
dimension	String	Subscale the item belongs to (e.g. BFI: openness/conscientiousness/extraversion/agreeableness/neuroticism; IOS: emotional_support/attention/positive_stimulation/social_comparison; STAS: trust_in_automation).
orientation	String (direct / inverted)	Whether the item is reverse-coded when computing the dimension score.

final_survey_answers.csv, intermediate_survey_answers.csv

AATL questionnaire answers: final_survey_answers.csv holds the 53-item questionnaire answered at the end of each block, and intermediate_survey_answers.csv the shorter 4-factor questionnaire answered after each trial.

Column	Type	Description
session_id	Integer	Subject/session identification.
question_id	Integer	Item identification.
session_block_id	Integer (final only)	Block the answer refers to; matches `session_block_id` in `metadata.csv`.
session_trial_id	Integer (intermediate only)	Trial the answer refers to; matches `session_trial_id` in `metadata.csv`.
name	String	final or intermediate.
text, text_english	String	Item text, in Spanish and English (`[Agent Name]` refers to the agent being evaluated).
response	Integer (1-5)	Subject's answer, 1-5 Likert scale.
dimension	String	AATL subscale: autonomy, altruism, cooperative, relationship-building, communication, synchronized, ability (final); altruism, communication, relationship-building, ability (intermediate).
orientation	String (direct / inverted)	Whether the item is reverse-coded when computing the dimension score.

neutral_silero_vad_aligns.csv

Voice-activity alignment (Silero VAD) for the neutral speech recordings (neutral_reading.wav, neutral_free_speech.wav).

Column	Type	Description
session_id	Integer	Subject/session identification.
task_type	String (neutral_reading / neutral_free_speech)	Which of the two neutral recordings this segment belongs to.
start, end	Float (seconds)	Segment boundaries within the recording.
dataset_path	String	Path to the corresponding audio file, relative to dataset/ (e.g. active-priming/audios/session_289/neutral_free_speech.wav).

silero_vad_aligns.csv

Voice-activity alignment (Silero VAD) for the game recordings (block_<1-2>-trial_<1-4>-stage_<1-2>.wav).

Column	Type	Description
start, end	Float (seconds)	Segment boundaries within the recording.
dataset_path	String	Path to the corresponding stage recording, relative to dataset/, of the form active-priming/audios/session_<session_id>/block_<block_order>-trial_<trial_order>-stage_<stage>.wav.

Dataset collection

• Inclusion criteria: age 18-65, Rioplatense Spanish accent, normal or corrected-to-normal visual acuity and color vision.
• Recruitment: laboratory's social media, student mailing lists, posters on campus.
• Ethics: approved by the Institutional Review Board of Hospital General de Agudos J. M. Ramos Mejía (IRB No. 00001942; OHRP FWA 00001767; protocol registration code 6215; approval date November 30, 2021), under an exemption per 45 CFR §46.104. Subjects gave prospective consent after a general description of the study; full protocol details were disclosed by email after the experiment.
• Sessions were remote, using the subject's own computer, microphone and headset.
• Some sessions were interrupted (e.g. connection issues) and resumed later; these are included in the dataset (flagged via what_discontinuous_task/where_discontinuous_task in metadata.csv) but excluded from the analyses.

Publication

The study associated with this dataset is currently in the process of publication. A citation will be added here once it is out.