Adamantium — Multilingual Golden Sample ASR Dataset

A curated, multilingual speech dataset for evaluating and fine-tuning ASR models (e.g., OpenAI Whisper).

27,603+ audio clips across 4 languages with consistent formatting, balanced splits, and rich metadata.

Dataset Summary

• Total samples: 27,603
• Languages: 4 (Malay, Tamil, English, Mandarin Chinese)
• Splits: Train (21,968) / Validation (2,767) / Test (2,868)
• Total duration: ~42.16 hours
• Format: HuggingFace Dataset + Parquet (self-contained, ~6.6GB)
• Audio spec: WAV PCM 16-bit, 16 kHz mono, 3–30 sec per clip
• Audio: Fully embedded in parquet files - no separate downloads needed

Languages

• Malay (ms): 12,753 samples across 5 dialects (standard, kelantan, sabah-bisaya, sarawak-kelambit, sarawak-serian-bidayuh)
• Tamil (ta): 4,284 samples (OpenSLR SLR65)
• English (en): 5,567 samples (LibriSpeech)
• Mandarin Chinese (zh): 4,999 samples (MAGICDATA)

Features

Each sample includes:

• file_name: Path to audio file
• transcription: Ground-truth text
• language: ISO 639-1 code
• dialect: Regional variant
• speaker_id: Unique speaker identifier (when available)
• gender: M, F, or unknown
• age_group: teen, young_adult, adult, senior (when available)
• duration: Audio length in seconds
• sample_rate: 16000 Hz
• domain: conversational or scripted
• code_switch: Whether sample contains code-mixing
• audio_bytes: Raw audio bytes (WAV PCM16, 16 kHz mono) - fully embedded in parquet

Loading

from datasets import load_dataset

# Load from local parquet files
ds = load_dataset("parquet", data_files={
    "train": "train-*.parquet",
    "validation": "validation-*.parquet",
    "test": "test-*.parquet"
})

# Or load from HuggingFace Hub (once published)
ds = load_dataset("radii/adamantium")

# Access specific split
train = ds["train"]

Usage Example

from datasets import load_dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import soundfile as sf
import io

# Load dataset
ds = load_dataset("parquet", data_files="train-*.parquet", split="train")

# Load Whisper model
processor = WhisperProcessor.from_pretrained("openai/whisper-base")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")

# Prepare sample
sample = ds[0]
# Decode audio bytes to numpy array
audio_array, sampling_rate = sf.read(io.BytesIO(sample["audio_bytes"]))

# Process
input_features = processor(audio_array, sampling_rate=sampling_rate, return_tensors="pt").input_features
predicted_ids = model.generate(input_features)

# Decode
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(f"Transcription: {transcription}")
print(f"Ground truth: {sample['transcription']}")

Validation

• All samples: audio format validated (16 kHz mono)
• Sample rate: 16000 Hz (consistent)
• Duration: 3–30 seconds (with edge cases <1s in corpus data)
• Metadata: Non-empty transcriptions, valid language codes

Citation

Please cite individual source corpora:

• Tamil: CC BY-SA 4.0 — He et al. (2020). LREC 2020.
• Mandarin: CC BY-NC-ND 4.0 — MAGICDATA (Magic Data Technology Co., Ltd., 2019)
• English: CC BY 4.0 — LibriSpeech (Panayotov et al., 2015, ICASSP)
• Malay: Mixed licenses (HF datasets + local corpus)

Version History

• v1.1 (2026-06-18): Malay dialect expansion (Sabah Bisaya, Sarawak variants) — 27,603 samples
• v1.0 (2026-06-15): Initial release with 4 languages — 22,756 samples

---

For detailed information about data sources and transformation pipeline, see the source repository documentation.