StyleInstructCapsDB / README.md
StyleInstructCaps's picture
Add files using upload-large-folder tool
8b2bea8 verified
|
Raw History Blame
9.02 kB
metadata
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - text-to-speech
configs:
  - config_name: StyleInstructCaps
    data_files:
      - split: train
        path: data/StyleInstructCaps/train-*.parquet
      - split: eval
        path: data/StyleInstructCaps/eval-*.parquet
  - config_name: StyleInstructCaps-MetaSet
    data_files:
      - split: test
        path: data/StyleInstructCaps-MetaSet/test-*.parquet
  - config_name: In_domain_balanced_set
    data_files:
      - split: test
        path: data/In_domain_balanced_set/test-*.parquet
  - config_name: VCTK_speaker_prompts
    data_files:
      - split: test
        path: data/VCTK_speaker_prompts/test-*.parquet
  - config_name: VCTK_unseen_prompts
    data_files:
      - split: test
        path: data/VCTK_unseen_prompts/test-*.parquet
  - config_name: Gigaspeech_SpeechCraft_captions
    data_files:
      - split: test
        path: data/Gigaspeech_SpeechCraft_captions/test-*.parquet

StyleInstructCaps

Style-captioned speech metadata for speaking-style captioning and instruction-following evaluation, packaged in a CapSpeech-style Hugging Face dataset layout.

Repository placeholder: <HF_ORG>/<DATASET_NAME>

Overview

This dataset release contains speaking-style captions, caption-generation instructions (when available), and minimal utterance metadata for speech captioning research. Each config includes only:

  • source, relative_audio_path, transcription, speakerid, duration (when present in the source manifest)
  • Speaking-style caption fields (when present)
  • Instruction fields for caption generation (when present)

Configs included:

  • StyleInstructCaps — large-scale training and evaluation splits
  • StyleInstructCaps-MetaSet — Eval-500 benchmark subset (434 utterances)
  • In-domain balanced set — accent/gender-balanced in-distribution evaluation set
  • VCTK speaker prompts — VCTK evaluation with 10 speaker-idiosyncratic style prompts
  • VCTK (unseen prompts) — VCTK out-of-distribution prompt evaluation (32 utterances, 36 prompts per row)
  • Gigaspeech (with SpeechCraft captions) — unseen GigaSpeech evaluation set with SpeechCraft reference captions

This is a metadata-only release. Audio is not redistributed. Each row includes a relative_audio_path that should be resolved against the corresponding upstream audio corpus listed below.

During release preparation, local absolute audio_path values are removed, mixed string/list caption fields are normalized to consistent types, and duration is stored as float64.

License

All resources are released under the CC BY-NC 4.0 license.

Usage

from datasets import load_dataset

REPO = "<HF_ORG>/<DATASET_NAME>"

train = load_dataset(REPO, "StyleInstructCaps", split="train")
eval_split = load_dataset(REPO, "StyleInstructCaps", split="eval")
metaset = load_dataset(REPO, "StyleInstructCaps-MetaSet", split="test")
in_domain = load_dataset(REPO, "In_domain_balanced_set", split="test")
vctk_speaker = load_dataset(REPO, "VCTK_speaker_prompts", split="test")
vctk_unseen = load_dataset(REPO, "VCTK_unseen_prompts", split="test")
gigaspeech = load_dataset(REPO, "Gigaspeech_SpeechCraft_captions", split="test")

print(train[0])

Dataset Structure

Each config is stored as parquet files under data/<config>/. Tags, phonemes, SNR, gender/accent labels, and other non-caption metadata from the source manifests are not included in this release.

Shared release columns

These columns appear in a config when present in the source manifest:

Column Type Description
source string Upstream dataset identifier
relative_audio_path string Relative path to the audio file within the upstream corpus
transcription string Transcript of the utterance
speakerid string Speaker identifier
duration float Utterance duration in seconds

Style caption and instruction columns are config-specific; see below.

StyleInstructCaps

Split Rows Description
train 1,031,840 Full-scale training set (11 parquet shards)
eval 14,756 Held-out evaluation set with instructions

Style caption columns (train and eval):

Column Type Description
text_description list[string] Natural-language style caption(s)
speaker_idiosyncratic_style string Speaker identity style caption
situational_contextual_style string Situational/contextual style caption
expressive_emotional_style string Expressive/emotional style caption
linguistic_pragmatic_style string Linguistic/pragmatic style caption
perceptual_listener_centric_style string Listener-centric style caption
holistic_creative_synthesis string Holistic creative synthesis caption

Instruction columns (eval only):

Column Type Description
speaker_idiosyncratic_style_instruction string Instruction for speaker style caption
situational_contextual_style_instruction string Instruction for situational style
expressive_emotional_style_instruction string Instruction for expressive style
linguistic_pragmatic_style_instruction string Instruction for linguistic style
perceptual_listener_centric_style_instruction string Instruction for listener-centric style
holistic_creative_synthesis_instruction string Instruction for holistic synthesis

StyleInstructCaps-MetaSet

Split Rows Description
test 434 Eval-500 benchmark subset

Schema matches StyleInstructCaps/eval (base columns, all style caption fields, and all six *_instruction columns).

In-domain balanced set

Split Rows Description
test 400 Accent/gender-balanced in-distribution evaluation set

Columns: base metadata (source, relative_audio_path, speakerid, duration) plus prompt1_speaker_idiosyncratic_style_instruction through prompt10_speaker_idiosyncratic_style_instruction.

VCTK speaker prompts

Split Rows Description
test 320 VCTK evaluation with 10 speaker-idiosyncratic style prompts

Columns: base metadata (source, relative_audio_path, speakerid) plus prompt1_speaker_idiosyncratic_style_instruction through prompt10_speaker_idiosyncratic_style_instruction.

VCTK (unseen prompts)

Split Rows Description
test 32 VCTK OOD prompt evaluation (36 instruction prompts per utterance)

Columns: base metadata (source, relative_audio_path, speakerid) plus prompt1_speaker_idiosyncratic_style_instruction through prompt36_holistic_creative_synthesis_style_instruction.

Gigaspeech (with SpeechCraft captions)

Split Rows Description
test 600 Unseen GigaSpeech evaluation set

Columns:

Column Type Description
source string Upstream dataset identifier (eval_giga)
relative_audio_path string Relative path within the evaluation GigaSpeech subset
ground_truth_caption string SpeechCraft reference style caption
holistic_creative_synthesis_instruction string Instruction for holistic caption generation

Audio Sources

Map each row's source and relative_audio_path to the corresponding upstream corpus:

source value Upstream audio corpus
emilia Emilia-Dataset (EN)
voxceleb VoxCeleb / VoxCeleb2
VCTK VCTK Corpus
ears EARS Dataset
expresso Expresso
eval_giga GigaSpeech (evaluation subset)

Preparing and Uploading (maintainers)

From the project root:

python -m venv .venv
source .venv/bin/activate
pip install -r release/scripts/requirements.txt

python release/scripts/prepare_release.py

# Edit release/README.md: replace <HF_ORG>/<DATASET_NAME> and citation
huggingface-cli login
python release/scripts/upload_to_hf.py --org <HF_ORG> --name <DATASET_NAME>

Citation

@misc{styleinstructcaps2026,
  title        = {StyleInstructCaps: Style-Captioned Speech Annotations},
  author       = {<TODO: authors>},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/<HF_ORG>/<DATASET_NAME>}}
}

Replace <HF_ORG>, <DATASET_NAME>, and the citation fields before publishing.