Panhapich/fastmss_meeting_simulation

FastMSS Khmer Meeting Simulation

This dataset contains 6,900 simulated Khmer multi-speaker meetings (130.37 hours) generated with FastMSS from the Digital-Divide-Data/khmer-speech-dataset speech corpus.

The meetings are intended for speaker diarization research, especially for experimenting with Khmer data when real multi-speaker meeting recordings are not available.

Important: these are synthetic, clean, read-speech meetings, not real conversational recordings. They contain only 20 source speakers and were generated without reverberation, background noise, or FIR filtering. Results on this dataset should not be treated as evidence of real-world performance.

At a glance

Item	Value
Language	Khmer (km)
Meetings	6,900
Audio	130.37 hours
Speakers	20
Speakers per meeting	2, 3, or 4
Topics	69
Sample rate	16 kHz
Channels	Mono
Audio format	WAV, 16-bit PCM
Annotation	RTTM + JSONL metadata
Per-speaker stems	Not included

Source data and generation

The source speech is the Cultural Khmer ASR Dataset:

• Dataset: https://huggingface.co/datasets/Digital-Divide-Data/khmer-speech-dataset
• License: CC BY-SA 4.0

Meetings were generated with FastMSS:

• Code: https://github.com/popcornell/FastMSS
• Recipe: FastMSS/recipes/khmer.yaml
• Source manifests: one Lhotse manifest per topic, then combined for simulation

The simulation used these settings:

• 100 meetings per topic, 69 topics
• 2–4 speakers per meeting, at most 4 concurrent speakers
• 60 s target duration; actual meetings are 60–318 s because turn-taking extends the target
• 16 kHz mono output
• reverberate: false
• add_noise: false
• use_fir: false
• save_rttm: true
• Flat turn-taking priors, with interruptions and backchannels producing overlap

RTTMs were generated from the simulation's aligned speech segments. Same-speaker segments separated by 0.2 s or less were merged. This is why the number of RTTM segments differs from the number of utterance-level manifest segments.

See ATTRIBUTION.md for source and tool attribution.

Splits

The split is a designated-speaker holdout:

• Train: 6,207 meetings / 117.18 h / 16 speakers
• Dev: 611 meetings / 11.50 h / 17 speakers
• Test: 82 meetings / 1.70 h / 16 speakers

No designated dev or test speaker appears in a training meeting. However, the underlying speech cuts are reused across topics and splits:

• 451 source cuts overlap train and dev
• 39 source cuts overlap train and test

This is not a fully source-cut-disjoint split. Keep that in mind when interpreting very high scores.

Split	Meetings	Hours	Speakers in meetings	Source cuts	Manifest segments	RTTM segments	2-spk	3-spk	4-spk
train	6,207	117.176	16	49,542	58,556	50,972	2,105	2,149	1,953
dev	611	11.497	17	5,511	6,053	5,318	127	228	256
test	82	1.699	16	604	715	604	17	34	31
total	6,900	130.373	20	55,151	65,324	56,894	2,249	2,411	2,240

Overlap statistics are provided in metadata/stats.json:

• overlap_ratio = overlap seconds / union of speech seconds
• overlap_fraction = overlap seconds / meeting duration

Dataset-wide mean overlap_ratio is 0.2023 (p95 0.4038); mean overlap_fraction is 0.1955 (p95 0.3919).

Files

audio/
  <meeting_id>.wav
rttm/
  train/<meeting_id>.rttm
  dev/<meeting_id>.rttm
  test/<meeting_id>.rttm
metadata/
  train.jsonl
  dev.jsonl
  test.jsonl
  stats.json
  split.json
  source_validation.json

Every WAV is a mono 16 kHz meeting mixture. Every RTTM uses the standard SPEAKER format with speaker labels such as f-adt1-0001 and m-adt2-0002.

Metadata schema

Each JSONL line describes one meeting:

{
  "meeting_id": "00150280-3181-4e63-b7e1-6aae3e7ffe45",
  "split": "train",
  "audio": "audio/00150280-3181-4e63-b7e1-6aae3e7ffe45.wav",
  "rttm": "rttm/train/00150280-3181-4e63-b7e1-6aae3e7ffe45.rttm",
  "duration": 60.0,
  "sampling_rate": 16000,
  "channels": 1,
  "topic": "wildlife-condition-in-cambodia",
  "num_speakers": 3,
  "speakers": ["f-adt2-0002", "m-adt1-0004", "m-adt2-0002"],
  "num_segments": 9,
  "num_rttm_segments": 8,
  "speech_seconds": 58.0,
  "overlap_seconds": 5.0,
  "overlap_ratio": 0.086,
  "overlap_fraction": 0.083,
  "segments": [
    {
      "start": 0.9115,
      "end": 7.9415,
      "speaker": "m-adt2-0002",
      "text": "...",
      "transition_type": "FIRST",
      "source_cut_id": "train-...",
      "speech_level_db": -22.97
    }
  ]
}

segments are utterance-level supervisions from the FastMSS manifest and include the Khmer transcript and source-cut metadata. The RTTM is the diarization reference and may merge adjacent same-speaker segments. Use the RTTM for diarization scoring; use segments when you need transcripts or source-cut provenance.

Loading

The JSONL stores paths relative to the dataset root. Download a snapshot and resolve them locally:

import json
from pathlib import Path

from huggingface_hub import snapshot_download

root = Path(
    snapshot_download(
        "Panhapich/fastmss_meeting_simulation",
        repo_type="dataset",
    )
)

records = [
    json.loads(line)
    for line in (root / "metadata" / "train.jsonl").open(encoding="utf-8")
]
meeting = records[0]
wav_path = root / meeting["audio"]
rttm_path = root / meeting["rttm"]

With datasets, you can load the metadata directly:

from datasets import load_dataset

ds = load_dataset(
    "Panhapich/fastmss_meeting_simulation",
    data_files={
        "train": "metadata/train.jsonl",
        "dev": "metadata/dev.jsonl",
        "test": "metadata/test.jsonl",
    },
)

The audio and rttm columns remain relative paths; prefix them with the local dataset root before opening the files.

Intended use and limitations

Suitable for:

• Khmer speaker diarization experiments
• Pipeline and scoring smoke tests
• Comparing diarization systems under controlled overlap
• Studying the effect of simulated conversational data

Not suitable for:

• Claims about real Khmer meeting performance
• Commercial deployment without a separate license review
• Speaker extraction or separation benchmarks requiring per-speaker references
• Evaluating robustness to noise, reverberation, far-field microphones, or spontaneous conversational speech

Main limitations:

• Only 20 source speakers
• Clean, close-talk, read speech
• No noise, reverberation, or FIR filtering
• Source cuts overlap across splits
• No per-speaker stems
• Synthetic turn-taking may not match real meeting dynamics

License and attribution

The dataset is released under CC BY-SA 4.0, following the source corpus license. See LICENSE and ATTRIBUTION.md.

If you use this dataset, please also cite FastMSS:

@unpublished{polok2026mindthegap,
  author = {Polok, Alexander and Medennikov, Ivan and Watanabe, Shinji and
            {\v{C}}ernock{\'y}, Jan and Burget, Luk{\'a}{\v{s}} and
            Cornell, Samuele},
  title = {Mind the Gap: Impact of Synthetic Conversational Data on
           Multi-Talker {ASR} and Speaker Diarization},
  note = {Submitted to INTERSPEECH},
  year = {2026}
}