aisingapore/seahelm-audio

Dataset Card for SEA-HELM Audio (OmniVoice Baseline)

[Last Update: 21 Oct 2026]

Dataset Summary

SEA-HELM Audio is a Southeast Asian multilingual voice benchmark designed to quantify the performance degradation introduced by the speech modality in end-to-end conversational voice stacks across Southeast Asian languages.

Adapted from AI Singapore's text-based SEA-HELM (Southeast Asian Holistic Evaluation of Language Models) benchmark, this dataset provides a standardized input speech corpus generated exclusively using OmniVoice. By pairing synthesized input audio with ground-truth prompt text, the dataset enables dual-branch evaluation to isolate the ASR modality gap

$$\mathrm{\Delta }(ModalityGap)={\text{Score}}_{\text{text}}-{\text{Score}}_{\text{speech}}\backslash Delta\; (Modality\; Gap)\; =\; \backslash text\{Score\}\_\{\backslash text\{text\}\}\; -\; \backslash text\{Score\}\_\{\backslash text\{speech\}\}$$

under deterministic, rule-based verification.

Supported Tasks and Languages

The corpus covers two primary evaluation tasks across five target languages:

Task Subset	Task Type	Languages Covered	Evaluation Protocol
seahelm_ifeval	Instruction Following	English (en), Indonesian (id), Malay (ms), Tamil (ta)	Rule-based constraint verifier + language compliance check
seahelm_global_mmlu_lite	Multiple-Choice QA	Chinese (zh), English (en)	Deterministic option letter extraction (A–D)

Language Breakdown

• Chinese (zh): Simplified Chinese
• English (en): Regional and general English
• Indonesian (id): Bahasa Indonesia
• Malay (ms): Bahasa Melayu
• Tamil (ta): Regional Tamil

Data Generation Methodology

All input audio in this public release was generated using OmniVoice to establish a unified, open-weight acoustic baseline].

1. Acoustic Standardization: Prompts are read aloud using OmniVoice to simulate standardized user acoustic inputs across all regional target languages.
2. Linguistic Adaptation: Prompts are formatted to match natural spoken cadence. Instruction-following prompts (seahelm_ifeval) are spoken verbatim as written, while multiple-choice questions (seahelm_global_mmlu_lite) incorporate task preambles alongside question bodies and options.
3. Deterministic Answer Isolation: Ground-truth labels, constraint parameters, and answer keys are withheld from the audio signal and stored purely in the text sidecar for scoring time.

Dataset Structure

Data Instances

Each instance represents a single evaluation prompt, containing the ground-truth text, metadata, and the synthesized audio array.

An example from the dataset looks as follows:

{
  "sample_id": "ifeval_ms_001",
  "audio": {
    "path": "ifeval_ms_001.wav",
    "array": [0.001, 0.002, ...],
    "sampling_rate": 16000
  },
  "text": "Tuliskan sebuah karangan pendek. Pastikan ia kurang daripada 50 perkataan.",
  "language": "ms",
  "task_subset": "seahelm_ifeval"
}

Data Fields

• id (string): Unique sample identifier.
• language (string): ISO language code (zh, en, id, ms, ta).
• subset (string): SEA-HELM subtask (seahelm_ifeval or seahelm_global_mmlu_lite).
• audio (datasets.Audio): Audio object containing synthesized speech waveform (24 kHz)) generated via OmniVoice.
• prompt_text (string): The original ground-truth text prompt supplied to the text baseline branch.
• tts_engine (string): Input acoustic synthesizer used (OmniVoice).
• target_label / constraints (dict/string): Ground-truth target key used by the programmatic verification engine.

Annotations

No human annotations were added. The dataset relies on the ground-truth labels and rule-based verification scripts provided by the parent SEA-HELM benchmark.

Considerations for Using the Data

Social Impact and Limitations

• Synthetic Acoustic Distribution: The audio in this dataset is entirely generated via OmniVoice. While this provides a highly controlled, reproducible environment for isolating ASR-LLM degradation, it does not perfectly replicate the acoustic noise, hesitation markers, and varied microphone quality of authentic human speech in real-world deployments.
• Dialect Representation: Users should note that the TTS models may impose specific dialectal or prosodic assumptions on the generated regional languages (e.g., specific accents for Tamil or Malay) that might not encompass all local variations.

Quickstart Usage

You can load the dataset directly via the Hugging Face datasets library:

from datasets import load_dataset

# Load the full SEA-HELM Audio benchmark
dataset = load_dataset("aisingapore/seahelm-audio")

# Load a specific language and subset
ifeval_ms = load_dataset(
    "aisingapore/seahelm-audio", 
    name="seahelm_ifeval", 
    split="malay"
)

# Inspect a single sample
sample = ifeval_ms[0]
print("Prompt:", sample["prompt_text"])
print("Audio Array Shape:", sample["audio"]["array"].shape)

Evaluation Protocol

When evaluating a voice stack on this dataset, compare two parallel runs:

1. Text Branch: Supply prompt_text directly to your LLM to obtain Score_text.

2. Speech Branch: Pass the audio signal through your ASR model, then feed the resulting transcript to your LLM to obtain Score_speech.

Repository: aisingapore/seahelm-audio

Technical Report and Citation

If you utilize this corpus or our methodology in your work, please cite the corresponding paper:

@article{seahelm_audio_bench_en_zh_ta_ms_2026,
    title={SEA-HELM Audio: Evaluating Multilingual ASR-LLM Pipeline Performance Across SEA-HELM Subsets}, 
    author={Aulia Adila and Yixian Tan and Jonathan Heng and Kunat Pipatanakul and Donghang Wu and Anocha Sutaveephamochanon},
    year={2026},
    eprint={2610.19869},
    archivePrefix={arXiv},
    primaryClass={cs.SD},
    url={https://arxiv.org/abs/2610.19869}, 
}

Team

AI Products Pillar, AI Singapore

Acknowledgement

This project is supported by the National Research Foundation Singapore and Infocomm Media Development Authority (IMDA), Singapore under its National Large Language Model Funding Initiative.

Contact

sealion@aisingapore.org