HaessigDB

HaessigDB (human-annotated emotional speech snippets with intensity grading) is a corpus of acted, call-center-style English speech for fine-grained irritability detection. Four professional voice actors recorded synthetic banking-support dialogues scripted so the customer grows increasingly irritable as the request goes unresolved.

Each snippet carries crowdsourced ordinal intensity ratings (1–10) on three irritability dimensions: annoyance, frustration, and aggression. Unlike categorical resources such as Emo-DB, IEMOCAP, or RAVDESS, the labels capture how intense an emotion is rather than only its presence. Snippets retain their position within each call, which supports trajectory modeling and early-escalation research.

Subsets

The CSV subsets below are browsable in the Dataset Viewer. Each row links to an audio clip via the file_name column (see Audio files).

Subset	Contents
all_ratings	Full rated set with all three dimensions
aggression_high_agreement	High-agreement subset (Krippendorff's α > 0.80) for aggression
annoyance_high_agreement	High-agreement subset for annoyance
frustration_high_agreement	High-agreement subset for frustration
inner_join_high_agreement	Snippets meeting α > 0.80 across all three dimensions
outer_join_high_agreement	Snippets meeting α > 0.80 in at least one dimension

Audio files

The audio is stored as WAV files in the audio/ folder at the repository root. It is not embedded in the CSV subsets, so the clips do not play inside the Dataset Viewer. Each file is named exactly as the file_name value in the CSVs, so a row with file_name = actor1_call_1_sentence_1.wav corresponds to audio/actor1_call_1_sentence_1.wav.

To download a single clip referenced by a row:

from huggingface_hub import hf_hub_download

wav_path = hf_hub_download(
    repo_id="nwllr/haessigDB",
    filename="audio/actor1_call_1_sentence_1.wav",
    repo_type="dataset",
)

To download the full audio folder at once:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="nwllr/haessigDB",
    repo_type="dataset",
    allow_patterns="audio/*",
)

Fields

Column	Description
file_name	Audio filename, encoding actor, call, and sentence position (e.g. actor1_call_1_sentence_1.wav). The clip itself is at audio/<file_name>
actor	Voice actor identifier (group on this for speaker-disjoint splits)
call	Call identifier
sentence	Snippet position within the call (temporal index)
transcript	Text spoken in the snippet
aggression / frustration / annoyance	Mean ordinal intensity rating across annotators (1–10)

Intended uses

Supervised intensity prediction, emotion-trajectory modeling, and the design of escalation-aware conversational systems.

Loading

from datasets import load_dataset

# Load a label subset (browsable in the viewer)
ds = load_dataset("nwllr/haessigDB", "aggression_high_agreement")

# Pair a row with its audio file
from huggingface_hub import hf_hub_download
row = ds["train"][0]
wav_path = hf_hub_download(
    repo_id="nwllr/haessigDB",
    filename=f"audio/{row['file_name']}",
    repo_type="dataset",
)

Citation

Weller, N., Grau, M., & Blohm, I. (2026). HaessigDB: A Database of Irritable Speech with Intensity Grading. Interspeech 2026.