---
language:
- km
license: other
task_categories:
- automatic-speech-recognition
pretty_name: Khmer ASR v4 (actableai)
size_categories:
- 100K<n<1M
tags:
- khmer
- asr
- speech
- audio
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: duration
    dtype: float32
  - name: source
    dtype: string
  splits:
  - name: train
    num_bytes: 56199107865
    num_examples: 486046
  - name: validation
    num_bytes: 1181228871
    num_examples: 9910
  download_size: 59169192403
  dataset_size: 57380336736
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---

# Khmer ASR Dataset (v4)

A consolidated Khmer (ខ្មែរ) speech corpus used to train the actableai Khmer ASR
model (NeMo, v4 checkpoint). Combines 17 sub-corpora spanning call-center
recordings, broadcast/cultural content, public ASR corpora, and TTS-synthesised
speech.

## Summary

| Split      | Utterances | Hours   |
|------------|-----------:|--------:|
| train      |    486,046 |  502.77 |
| validation |      9,910 |   10.20 |
| **total**  |  **495,956** | **512.97** |

- Sampling rate: 16 kHz, mono, PCM_16.
- Text: Khmer script (light normalization from the training pipeline only).

## Schema

| Field     | Type              | Description                                |
|-----------|-------------------|--------------------------------------------|
| audio     | `Audio(16000 Hz)` | Mono waveform embedded in parquet.         |
| text      | `string`          | Khmer transcript.                          |
| duration  | `float32`         | Seconds.                                   |
| source    | `string`          | Sub-corpus tag (see below).                |

## Composition by source

| Source               | Train     | Validation |
|----------------------|----------:|-----------:|
| metfone_v1_v3        |   173,836 |      3,547 |
| synth_full           |   107,488 |      2,193 |
| ddd_cultural         |    55,570 |      1,134 |
| sethisak_en_kh       |    26,698 |        544 |
| sethisak_kh_en_v2    |    26,698 |        544 |
| rinabuoy_train       |    24,833 |        506 |
| sethisak             |    21,798 |        444 |
| djsamseng_large      |    17,628 |        359 |
| km_corpus            |    14,645 |        298 |
| sethisak_asr         |     5,545 |        113 |
| openslr42            |     2,848 |         58 |
| kheng                |     2,771 |         56 |
| mpwt                 |     2,017 |         41 |
| shunya               |     1,630 |         33 |
| moses                |       949 |         19 |
| rinabuoy_test        |       750 |         15 |
| grkpp                |       342 |          6 |

`synth_full` is TTS-synthesised speech produced by the actableai OmniVoice
pipeline — added in v4 to widen phonetic and prosodic coverage.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("actableai/data-khmer", split="train", streaming=True)
example = next(iter(ds))
print(example["text"], example["duration"], example["source"])
print(example["audio"]["sampling_rate"], example["audio"]["array"].shape)
```

For ASR evaluation, note that Khmer is a syllabic script without reliable
word boundaries — prefer **CER** over WER.

## Provenance & licensing

This is a research aggregation of multiple upstream sources. The licensing of
each sub-corpus follows its upstream project; downstream users are responsible
for verifying compliance with each source's terms before commercial use.
Synthesised audio (`synth_full`) is released under the same terms as the
upstream TTS voices it was produced from.