Tnaot/large-dataset-audio-v2

khmer_speech_dataset

Khmer speech dataset with transcriptions, speaker labels, and metadata.

Dataset Description

This dataset contains Khmer (Cambodian) speech recordings with detailed transcriptions and annotations.

Dataset Statistics

Metric	Value
Total Examples	9,285
Total Duration	336.68 hours
Average Duration	130.54 seconds
Total Words	1,336,611
Khmer Words	994,050 (74.4%)
English Words	340,127 (25.4%)
Unique Sources	2857
Unique Tags	2092

Source Distribution

Source Type	Count
youtube	5,947
unknown	16
conversation	696
telegram	2,626

Speaker Distribution

Speakers	Count	Duration (hours)
0	2	0.00 (0.0%)
1	3,060	33.97 (10.1%)
2	2,855	137.80 (40.9%)
3	2,892	141.56 (42.0%)
4	387	18.99 (5.6%)
5	65	3.18 (0.9%)
6	16	0.78 (0.2%)
7	2	0.10 (0.0%)
8	5	0.25 (0.1%)
12	1	0.05 (0.0%)

Top Tags

Tag	Count
overlapping	5,982
laughing	4,736
drag_tone	3,705
fast	3,337
mocking	1,821
high_pitch	1,792
shout	1,597
frustrated	1,163
smack_lips	1,161
heavy_breathing	961
whisper	886
softly	856
stutter	726
serious	664
normal	638
sigh	465
hesitant	433
normal_tone	428
panic	423
thinking	399

Dataset Structure

Data Fields

• audio: Audio file path/bytes
• text_raw: Raw transcription with speaker labels and tags
• text_clean: Cleaned transcription (no tags, no speaker labels)
• duration_seconds: Audio duration in seconds
• speaker_count: Number of speakers in the audio
• speakers: List of speaker IDs
• tags: List of normalized annotation tags
• english_words: Count of English words
• khmer_words: Count of Khmer words
• total_words: Total word count
• chunk_id: Chunk index within original audio
• original_name: Name of original audio file
• source_type: Source type (telegram, youtube, conversation, unknown)
• source_folder: Original folder name
• telegram_chat_id: Telegram chat ID (if applicable)

Tag Categories

• Speech markers: hesitant, thinking, confused, drag_tone, overlapping
• Audio markers: music, low_quality, noise
• Language markers: english, khmer

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("path/to/dataset")

# Access examples
for example in dataset["train"]:
    audio = example["audio"]
    text = example["text_clean"]
    print(f"Duration: {example['duration_seconds']:.2f}s")
    print(f"Text: {text[:100]}...")

License

This dataset is released under CC-BY-NC-4.0 license.

Citation

If you use this dataset, please cite:

@dataset{khmer_speech_dataset,
  title={Khmer Speech Dataset},
  year={2024},
  url={https://huggingface.co/datasets/...}
}