SaLTUNIMAS/sarawak-malay-asr

SARAWAK MALAY SPEECH DATASET

INTRODUCTION

This dataset contains speech and transcription data for Sarawak Malay, a low-resource dialect spoken in Sarawak, Malaysia. It was collected and prepared by the ASR research group at Universiti Malaysia Sarawak (UNIMAS) for Automatic Speech Recognition (ASR) experiments.

CORPUS OVERVIEW

Total utterances : 1,164
Total audio      : ~1.9 hours
Unique speakers  : 42
Unique words     : 2,329

SOURCES

SM1_ - Original Sarawak Malay Corpus : 738 utterances
       Long-form interviews recorded in natural conversation settings.
SM2_ - New Isolated Transcription    : 426 utterances
       Isolated sentence recordings by UNIMAS students.

SPEAKER INFORMATION

Gender is encoded in the utterance ID filename:
  SM_FF_ - Female interviewer, Female speaker
  SM_MM_ - Male interviewer, Male speaker
  SM_FM_ - Female interviewer, Male speaker
  SM_MF_ - Male interviewer, Female speaker

DATASET STRUCTURE

Columns in this dataset:
  audio         - WAV audio file at 16kHz
  transcription - Sarawak Malay transcript (original text, no cleaning)
  utterance_id  - Unique utterance identifier
  speaker_id    - Speaker name or topic from filename
  gender        - Speaker gender (male/female)
  source        - SM1_original_corpus or SM2_new_recording
  duration      - Audio segment duration in seconds
  split         - train or test

TRAIN/TEST SPLIT

Hard speaker-based split (80% train / 20% test), gender-balanced.
Split uses utterance-aware greedy assignment to maintain the 80/20
ratio separately for female and male speakers.
  train : ~930 utterances
  test  : ~234 utterances

AUDIO FORMAT

WAV files, mono, 16kHz sampling rate.
Utterance IDs follow the format:
  SM[1|2]_SM_[FF|FM|MF|MM]_SPEAKERNAME_SESSIONID_UTTERANCEID

ASR RESULTS (Fine-Tuned Wav2Vec2 XLS-R 300M)

Base model  : mesolitica/wav2vec2-xls-r-300m-mixed
  Greedy decoding (no LM) : 69.9% WER
  With KenLM (best a/b)   : 64.5% WER

HOW TO USE THIS DATASET

OPTION 1 - Load in Python (recommended)

Install the datasets library first:
  pip install datasets

Then load the dataset:
  from datasets import load_dataset
  ds = load_dataset("SaLTUNIMAS/sarawak-malay-asr")

Access a single sample:
  sample = ds['train'][0]
  print(sample['transcription'])   # the transcript
  print(sample['duration'])        # audio duration in seconds
  print(sample['speaker_id'])      # speaker name/topic name
  print(sample['gender'])          # male or female
  print(sample['source'])          # SM1 or SM2

Play audio in Jupyter or Google Colab:
  import IPython.display as ipd
  ipd.Audio(sample['audio']['array'],
            rate=sample['audio']['sampling_rate'])

Load only one specific split
  train_ds = load_dataset("SaLTUNIMAS/sarawak-malay-asr", split="train")
   ## OR ##
  test_ds  = load_dataset("SaLTUNIMAS/sarawak-malay-asr", split="test")

  ## OR ##
Load both split
  from datasets import concatenate_datasets
  all_data = concatenate_datasets([ds['train'], ds['test']])
  print(f"Total: {len(all_data)} utterances")

Filter by gender:
  female = ds['train'].filter(lambda x: x['gender'] == 'female')
  male   = ds['train'].filter(lambda x: x['gender'] == 'male')

Filter by source corpus:
  sm1 = ds['train'].filter(lambda x: x['source'] == 'SM1_original_corpus')
  sm2 = ds['train'].filter(lambda x: x['source'] == 'SM2_new_recording')

OPTION 2 - Save to disk (download once, use offline)

 from datasets import load_dataset, load_from_disk
 
 # Download and save
 ds = load_dataset("SaLTUNIMAS/sarawak-malay-asr")
 ds.save_to_disk("sarawak_malay_dataset")
 
 # Next time load from disk (no internet needed)
 ds = load_from_disk("sarawak_malay_dataset")

OPTION 3 - Download all files via Python

  from huggingface_hub import snapshot_download
  
  snapshot_download(
      repo_id="SaLTUNIMAS/sarawak-malay-asr",
      repo_type="dataset",
      local_dir="./sarawak_malay_dataset"
  )

OPTION 4 - Manual download

  Go to the Files and versions tab on this page and download
  the parquet files directly. Audio is embedded inside the
  parquet files so no separate download is needed.

ACKNOWLEDGEMENT

This dataset was collected and prepared at the Faculty of Computer Science
and Information Technology (FCSIT), Universiti Malaysia Sarawak (UNIMAS).
94300 Kota Samarahan, Sarawak, Malaysia.