COHP-ASR: A 356-Hour Khmer Speech Corpus from Oral History Interviews

Dataset Description

COHP-ASR is a large-scale Khmer automatic speech recognition (ASR) corpus derived from the Cambodian Oral History Project (COHP) at Brigham Young University. It comprises 640 interviews with Cambodian community members, totaling 356.6 hours of audio and 289,413 aligned speech segments.

The interviews are long-form oral histories recorded in natural conversational settings. Interviewees include survivors of the Khmer Rouge period discussing personal experiences and family histories. All recordings and transcripts were originally collected and published by COHP with permission for public release.

The corpus was constructed by applying an iterative forced-alignment refinement pipeline to existing manual transcripts, producing high-quality segment-level timestamps without additional human annotation. This two-stage alignment process (SRT-V1 -> SRT-V2) yielded a 9 percentage-point CER improvement over single-pass alignment.

Configurations

This dataset provides two configurations:

segmented (default)

Ready-to-train speech segments aligned to SRT-V2 timestamps. Each row contains a short audio clip and its corresponding Khmer transcription.

from datasets import load_dataset

ds = load_dataset("byumatrixlab/cohp-asr", "segmented", split="train")
print(ds[0])
# {'audio': {...}, 'transcription': 'ចា៎អ៊ុំ! សូមជំរាបសួរអ៊ុំ!', 'duration': 2.78, ...}

Columns:

Column	Type	Description
audio	Audio (16 kHz)	Segment audio waveform
transcription	string	Khmer text (NFC Unicode-normalized)
duration	float	Segment duration in seconds
ctc_confidence	float	CTC alignment confidence score (-1.0 if unavailable)
interview_id	string	Anonymized interview identifier (e.g., COHP_0001)

raw

Full-length, unsegmented interview audio with both SRT-V1 (bootstrap alignment) and SRT-V2 (refined alignment) transcripts embedded as text columns. Use this config if you want to re-run or improve on the alignment/segmentation pipeline. The raw interview audio files and SRT files are also available for direct download in the raw/audio/, raw/srt_v1/, and raw/srt_v2/ directories.

from datasets import load_dataset

ds = load_dataset("byumatrixlab/cohp-asr", "raw", split="train")

Columns:

Column	Type	Description
audio	Audio (16 kHz)	Full interview audio (embedded in Parquet)
interview_id	string	Anonymized interview identifier
srt_v1	string	Full SRT-V1 transcript text
srt_v2	string	Full SRT-V2 transcript text
audio_duration_sec	float	Total interview duration in seconds
num_segments_v2	int	Number of segments in SRT-V2
segment_duration_sec	float	Sum of segment durations in seconds

Dataset Statistics

Split	Interviews	Segments	Segment Hours
Train	620	275,696	327.9
Validation	10	7,072	7.5
Test	10	6,645	6.0
Total	640	289,413	341.4

Total interview audio duration is 356.6 hours. The segment duration total of 341.4 hours reflects only speech-containing segments; the difference of ~15 hours consists of inter-segment silences and gaps.

Audio format: 16 kHz, 16-bit PCM, mono WAV.

Splits are assigned deterministically by alphabetical order of interview identifiers: the first 620 interviews form the training set, the next 10 form validation, and the final 10 form the test set. This is fully deterministic and requires no random seed.

SRT Versions

• SRT-V1 (bootstrap): Produced by forced alignment using an untuned facebook/mms-1b-all model. Usable but contains boundary drift and occasional misalignment due to domain mismatch.
• SRT-V2 (refined): Produced by re-running forced alignment with an MMS model fine-tuned on SRT-V1 segments. The domain-adapted model produces more accurate alignment boundaries. Training on SRT-V2 yields a 9 percentage-point CER improvement over SRT-V1 with the same model architecture.

The segmented config uses SRT-V2 segments. The raw config includes both versions for researchers who wish to study or improve the alignment pipeline.

Evaluation Notes

• Use CER, not WER. Khmer script does not use spaces as word delimiters, making word-level tokenization ambiguous. Character Error Rate (CER) is the appropriate metric.
• Apply NFC Unicode normalization to both reference and hypothesis text before scoring. Khmer Unicode has multiple valid representations for some character sequences; NFC normalization ensures consistent comparison.
• Baseline results: MMS-CTC fine-tuned achieves 37.54% CER; Whisper Small fine-tuned achieves 40.82% CER on the test set. See the associated paper for full results.

Content Note

This dataset contains oral history interviews with Cambodian community members. Some interviews include firsthand accounts of the Khmer Rouge period (1975-1979), which may describe violence, loss, displacement, and trauma. While no access restrictions are required per COHP's public release policy, users should be aware that content may be emotionally difficult.

Citation

If you use this dataset, please cite the associated paper (accepted to IEEE SLT 2026, to appear):

@inproceedings{shurtz2026cohpasr,
  title     = {Iterative Forced Alignment for Low-Resource Oral
History: Building the COHP Khmer ASR Corpus},
  author    = {Shurtz, Ammon and Barrett, Thomas and Fosse, Erik and Sorenson, Lawry},
  booktitle = {Proc. IEEE Spoken Language Technology Workshop (SLT)},
  year      = {2026},
  note      = {Accepted; to appear},
}

Acknowledgments

This dataset is derived from recordings and transcripts collected by the Cambodian Oral History Project (COHP) at Brigham Young University. COHP has granted permission for public, non-gated release of this dataset. We gratefully acknowledge the COHP team and all interviewees who shared their stories.

License

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.