sulaimank/mentalhealth-asr

Luganda Mental Health Dialogues

This is a conversational speech dataset of simulated mental health counselling sessions in Luganda, recorded in Uganda. It is designed to support research in automatic speech recognition (ASR), speaker diarization, speaker role classification, and gender classification for low-resource African languages.

Each conversation features two speakers — a Helper (counsellor) and a Seeker (client) — discussing a mental health topic. Sessions were recorded in two environments: in-person and simulated phone calls. The dataset is annotated at the segment level with speaker role, gender, and Luganda transcript.

Dataset Structure

Each row corresponds to one speech segment and contains the following fields:

Column	Type	Description
audio	Audio	Segment audio at 16kHz
segment_id	string	Unique segment identifier
audio_id	string	Conversation identifier
start	float	Segment start time in original audio (seconds)
end	float	Segment end time in original audio (seconds)
duration	float	Segment duration (seconds)
transcript	string	Luganda transcription
role	string	Speaker role: Helper or Seeker
gender	string	Speaker gender: Male or Female
topic	string	Conversation topic
condition	string	Mental health condition discussed
environment	string	Recording environment: phone or in-person

Naming Conventions

Audio ID format:

{id}__{topic}_{condition}_{environment}

Example: 0013__isolation-loneliness_depression_in-person

Segment ID format:

{audio_id}__{RoleGender}__{segment_number}

Example: 0013__isolation-loneliness_depression_in-person__HF__001

Where H = Helper, S = Seeker, F = Female, M = Male.

Topics

Topic	Description
academic-pressure	Stress related to school or university
alcohol-and-drug-abuse	Substance use and addiction
cultural-stigma	Mental health stigma in cultural context
family-conflict	Family relationship difficulties
grief	Loss and bereavement
health-worries	Concerns about physical health
isolation-loneliness	Social withdrawal and loneliness
relationship	Romantic and interpersonal relationship issues
trauma	Past traumatic experiences
work-financial-pressure	Occupational and financial stress

Mental Health Conditions

Condition	Description
depression	Depressive symptoms
anxiety	Anxiety-related presentations
epilepsy	Epilepsy-related mental health concerns
psychosis	Psychotic presentations
schizophrenia	Schizophrenia-related presentations

Data Collection

Conversations were scripted and performed by trained speakers portraying a Helper (counsellor) and a Seeker (client). Sessions were recorded in two environments: in-person using a close-microphone setup, and simulated phone calls. Speaker pairs include same-gender (Male-Male, Female-Female) and mixed-gender (Male-Female) conversations.

Usage

from datasets import load_dataset

dataset = load_dataset("sulaimank/mental-health-luganda")

# Access a sample
sample = dataset["train"][0]
print(sample["transcript"])
print(sample["role"])
print(sample["gender"])
print(sample["topic"])

# Play audio
import IPython.display as ipd
ipd.Audio(sample["audio"]["array"], rate=sample["audio"]["sampling_rate"])

Intended Uses

• Luganda automatic speech recognition (ASR)
• Speaker diarization for conversational speech
• Speaker role classification (Helper vs. Seeker)
• Audio gender classification
• Low-resource African language speech research
• Mental health NLP research

Out-of-Scope Uses

• Clinical diagnosis or mental health assessment
• This dataset contains simulated conversations only — not real patient data
• Direct deployment in mental health applications without further evaluation

Limitations

• All conversations are simulated and scripted; they do not reflect real clinical interactions
• Luganda-English code-switching is present in some segments but not explicitly annotated
• The dataset covers a limited number of speakers and may not capture all dialectal variation in Luganda
• Gender distribution is not perfectly balanced across all splits