semantic-vad-eot / README.md
tchiayan's picture
Update README.md
7d82fa8 verified
|
Raw History Blame Contribute Delete
7.67 kB
metadata
dataset_info:
  - config_name: all
    features:
      - name: id
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: language
        dtype: string
      - name: duration
        dtype: float64
      - name: silence_spans
        list:
          - name: start
            dtype: float64
          - name: end
            dtype: float64
      - name: words
        list:
          - name: word
            dtype: string
          - name: start
            dtype: float64
          - name: end
            dtype: float64
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: ds_name
        dtype: string
    splits:
      - name: test
        num_bytes: 44364540
        num_examples: 1800
    download_size: 44360375
    dataset_size: 44364540
  - config_name: default
    features:
      - name: id
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: language
        dtype: string
      - name: duration
        dtype: float64
      - name: silence_spans
        list:
          - name: start
            dtype: float64
          - name: end
            dtype: float64
      - name: words
        list:
          - name: word
            dtype: string
          - name: start
            dtype: float64
          - name: end
            dtype: float64
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: ds_name
        dtype: string
    splits:
      - name: test
        num_bytes: 11588348
        num_examples: 400
    download_size: 11589518
    dataset_size: 11588348
  - config_name: ms
    features:
      - name: id
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: language
        dtype: string
      - name: duration
        dtype: float64
      - name: silence_spans
        list:
          - name: start
            dtype: float64
          - name: end
            dtype: float64
      - name: words
        list:
          - name: word
            dtype: string
          - name: start
            dtype: float64
          - name: end
            dtype: float64
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: ds_name
        dtype: string
    splits:
      - name: test
        num_bytes: 13837377
        num_examples: 500
    download_size: 13834631
    dataset_size: 13837377
configs:
  - config_name: all
    data_files:
      - split: test
        path: all/test-*
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
  - config_name: ms
    data_files:
      - split: test
        path: ms/test-*

Semantic VAD / EoT — Sampled Test Set (tchiayan/semantic-vad-eot)

This dataset is a small, sampled derivative of Scicom-intl/semantic-vad-eot, created for quick, lightweight evaluation and benchmarking of Semantic VAD / end-of-turn (EoT) detection models without needing to stream or download the full source dataset.

Only the test split of the source dataset is used. For each subset (language / domain), 100 examples are randomly sampled from a shuffle buffer, so this is intended for fast iteration and sanity-checking rather than as a full evaluation benchmark.

Configs (subsets)

This repo provides three configs, each containing a single test split assembled from different combinations of the source dataset's subsets:

Config Included source subsets Rows
default ms_dialects, ms_imda, ms_malaysian, ms_parliament 400
ms ms_dialects, ms_imda, ms_malaysian, ms_parliament, ms_science_english 500
all ms_dialects, ms_imda, ms_malaysian, ms_parliament, ms_science_english, en, fr, de, it, ja, ko, zh, pl, pt, ru, es, th, tr 1,800

default and ms focus on Malaysian speech (dialects, IMDA-derived, general Malaysian, parliamentary speech, and — for ms/all — Malaysian science/English content). all additionally covers the full set of languages available in the source dataset.

Load a specific config with:

from datasets import load_dataset

ds = load_dataset("tchiayan/semantic-vad-eot", "ms", split="test")

How this dataset was built

For every subset above, 100 rows were sampled from the source dataset's test split using streaming mode with a fixed shuffle buffer and seed, then tagged with their originating subset name (ds_name) and combined into a single test split per config. Audio was cast to 16 kHz.

from datasets import load_dataset

split_names = [
    "ms_dialects", "ms_imda", "ms_malaysian", "ms_parliament",
    "ms_science_english", "en", "fr", "de", "it", "ja", "ko", "zh",
    "pl", "pt", "ru", "es", "th", "tr",
]

ds_list = []
for split_name in split_names:
    ds = (
        load_dataset("Scicom-intl/semantic-vad-eot", split_name, split="test", streaming=True)
        .shuffle(buffer_size=2500, seed=42)
        .take(100)
        .map(lambda x: {"ds_name": split_name})
    )
    ds_list.append(ds)

from datasets import Dataset, Audio
ds = Dataset.from_generator(lambda: (item for ds in ds_list for item in ds))
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

The default and ms configs are simply subsets of the rows produced above, filtered to the relevant ds_name values, and all contains everything.

Because sampling uses a fixed seed (42) and buffer size (2500), the selection is reproducible against the source dataset, though results may shift if the upstream Scicom-intl/semantic-vad-eot dataset is updated.

Dataset structure

Each row corresponds to a single audio segment together with its transcript, word-level alignment, and labeled silence spans that mark candidate end-of-turn / pause locations. This mirrors the schema of the source dataset with one additional column (ds_name).

Column Type Description
id string Unique identifier for the example (source dataset ID).
audio audio Audio waveform, resampled to 16 kHz.
language string Language/locale code of the utterance (e.g. en, ms, zh).
duration float64 Duration of the audio clip in seconds.
silence_spans list[{start, end}] Labeled silence intervals within the clip, in seconds — used as candidate EoT / pause points.
words list[{word, start, end}] Word-level forced-alignment timestamps for the transcript.
messages list[{role, content}] Conversational-style transcript (typically a single user turn) corresponding to the audio.
ds_name string Name of the originating subset from Scicom-intl/semantic-vad-eot (e.g. ms_dialects, en).

Intended use

This dataset is meant for:

  • Quick local evaluation of Semantic VAD / end-of-turn detection models (e.g. against livekit/eot-bench-style pipelines) without pulling the full multi-million-row source dataset.
  • Sanity-checking model behavior across Malaysian dialects/domains (default, ms) or a broader multilingual mix (all).

It is not a substitute for evaluating on the full Scicom-intl/semantic-vad-eot test split, given the small per-subset sample size (100 rows).

Source dataset & license

This dataset is derived entirely from Scicom-intl/semantic-vad-eot and inherits its CC-BY-4.0 license. Please refer to the source dataset card for full details on data collection, forced alignment, and silence-span labeling methodology.

Citation

If you use this dataset, please cite the original source dataset:

@misc{scicom_semantic_vad_eot,
  title  = {semantic-vad-eot},
  author = {Scicom-intl},
  year   = {2025},
  url    = {https://huggingface.co/datasets/Scicom-intl/semantic-vad-eot}
}