StyleInstructCaps commited on
Commit
f441c0b
·
verified ·
1 Parent(s): 8b2bea8

Add files using upload-large-folder tool

Browse files
README.md CHANGED
@@ -35,22 +35,20 @@ configs:
35
 
36
  # StyleInstructCaps
37
 
38
- Style-captioned speech metadata for speaking-style captioning and instruction-following evaluation, packaged in a [CapSpeech](https://huggingface.co/datasets/OpenSound/CapSpeech)-style Hugging Face dataset layout.
39
-
40
- **Repository placeholder:** `<HF_ORG>/<DATASET_NAME>`
41
 
42
  ## Overview
43
 
44
  This dataset release contains speaking-style captions, caption-generation instructions (when available), and minimal utterance metadata for speech captioning research. Each config includes only:
45
 
46
- - `source`, `relative_audio_path`, `transcription`, `speakerid`, `duration` (when present in the source manifest)
47
  - Speaking-style caption fields (when present)
48
  - Instruction fields for caption generation (when present)
49
 
50
  Configs included:
51
 
52
  - **StyleInstructCaps** — large-scale training and evaluation splits
53
- - **StyleInstructCaps-MetaSet** — Eval-500 benchmark subset (434 utterances)
54
  - **In-domain balanced set** — accent/gender-balanced in-distribution evaluation set
55
  - **VCTK speaker prompts** — VCTK evaluation with 10 speaker-idiosyncratic style prompts
56
  - **VCTK (unseen prompts)** — VCTK out-of-distribution prompt evaluation (32 utterances, 36 prompts per row)
@@ -58,8 +56,6 @@ Configs included:
58
 
59
  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.
60
 
61
- 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`.
62
-
63
  ## License
64
 
65
  All resources are released under the [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license.
@@ -96,7 +92,6 @@ These columns appear in a config when present in the source manifest:
96
  | `relative_audio_path` | string | Relative path to the audio file within the upstream corpus |
97
  | `transcription` | string | Transcript of the utterance |
98
  | `speakerid` | string | Speaker identifier |
99
- | `duration` | float | Utterance duration in seconds |
100
 
101
  Style caption and instruction columns are config-specific; see below.
102
 
@@ -111,7 +106,6 @@ Style caption and instruction columns are config-specific; see below.
111
 
112
  | Column | Type | Description |
113
  | ------ | ---- | ----------- |
114
- | `text_description` | list[string] | Natural-language style caption(s) |
115
  | `speaker_idiosyncratic_style` | string | Speaker identity style caption |
116
  | `situational_contextual_style` | string | Situational/contextual style caption |
117
  | `expressive_emotional_style` | string | Expressive/emotional style caption |
@@ -134,7 +128,7 @@ Style caption and instruction columns are config-specific; see below.
134
 
135
  | Split | Rows | Description |
136
  | ----- | ---- | ----------- |
137
- | `test` | 434 | Eval-500 benchmark subset |
138
 
139
  Schema matches **StyleInstructCaps/eval** (base columns, all style caption fields, and all six `*_instruction` columns).
140
 
@@ -144,7 +138,7 @@ Schema matches **StyleInstructCaps/eval** (base columns, all style caption field
144
  | ----- | ---- | ----------- |
145
  | `test` | 400 | Accent/gender-balanced in-distribution evaluation set |
146
 
147
- **Columns:** base metadata (`source`, `relative_audio_path`, `speakerid`, `duration`) plus `prompt1_speaker_idiosyncratic_style_instruction` through `prompt10_speaker_idiosyncratic_style_instruction`.
148
 
149
  ### VCTK speaker prompts
150
 
@@ -189,32 +183,3 @@ Map each row's `source` and `relative_audio_path` to the corresponding upstream
189
  | `ears` | [EARS Dataset](https://github.com/facebookresearch/ears_dataset) |
190
  | `expresso` | [Expresso](https://github.com/facebookresearch/textlesslib/tree/main/examples/expresso/dataset) |
191
  | `eval_giga` | [GigaSpeech](https://huggingface.co/datasets/speechcolab/gigaspeech) (evaluation subset) |
192
-
193
- ## Preparing and Uploading (maintainers)
194
-
195
- From the project root:
196
-
197
- ```bash
198
- python -m venv .venv
199
- source .venv/bin/activate
200
- pip install -r release/scripts/requirements.txt
201
-
202
- python release/scripts/prepare_release.py
203
-
204
- # Edit release/README.md: replace <HF_ORG>/<DATASET_NAME> and citation
205
- huggingface-cli login
206
- python release/scripts/upload_to_hf.py --org <HF_ORG> --name <DATASET_NAME>
207
- ```
208
-
209
- ## Citation
210
-
211
- ```bibtex
212
- @misc{styleinstructcaps2026,
213
- title = {StyleInstructCaps: Style-Captioned Speech Annotations},
214
- author = {<TODO: authors>},
215
- year = {2026},
216
- howpublished = {\url{https://huggingface.co/datasets/<HF_ORG>/<DATASET_NAME>}}
217
- }
218
- ```
219
-
220
- Replace `<HF_ORG>`, `<DATASET_NAME>`, and the citation fields before publishing.
 
35
 
36
  # StyleInstructCaps
37
 
38
+ Style-captioned speech metadata for speaking-style captioning and instruction-following evaluation.
 
 
39
 
40
  ## Overview
41
 
42
  This dataset release contains speaking-style captions, caption-generation instructions (when available), and minimal utterance metadata for speech captioning research. Each config includes only:
43
 
44
+ - `source`, `relative_audio_path`, `transcription`, `speakerid` (when present in the source manifest)
45
  - Speaking-style caption fields (when present)
46
  - Instruction fields for caption generation (when present)
47
 
48
  Configs included:
49
 
50
  - **StyleInstructCaps** — large-scale training and evaluation splits
51
+ - **StyleInstructCaps-MetaSet** (434 utterances)
52
  - **In-domain balanced set** — accent/gender-balanced in-distribution evaluation set
53
  - **VCTK speaker prompts** — VCTK evaluation with 10 speaker-idiosyncratic style prompts
54
  - **VCTK (unseen prompts)** — VCTK out-of-distribution prompt evaluation (32 utterances, 36 prompts per row)
 
56
 
57
  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.
58
 
 
 
59
  ## License
60
 
61
  All resources are released under the [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license.
 
92
  | `relative_audio_path` | string | Relative path to the audio file within the upstream corpus |
93
  | `transcription` | string | Transcript of the utterance |
94
  | `speakerid` | string | Speaker identifier |
 
95
 
96
  Style caption and instruction columns are config-specific; see below.
97
 
 
106
 
107
  | Column | Type | Description |
108
  | ------ | ---- | ----------- |
 
109
  | `speaker_idiosyncratic_style` | string | Speaker identity style caption |
110
  | `situational_contextual_style` | string | Situational/contextual style caption |
111
  | `expressive_emotional_style` | string | Expressive/emotional style caption |
 
128
 
129
  | Split | Rows | Description |
130
  | ----- | ---- | ----------- |
131
+ | `test` | 434 | StyleInstructCaps-MetaSet evaluation subset |
132
 
133
  Schema matches **StyleInstructCaps/eval** (base columns, all style caption fields, and all six `*_instruction` columns).
134
 
 
138
  | ----- | ---- | ----------- |
139
  | `test` | 400 | Accent/gender-balanced in-distribution evaluation set |
140
 
141
+ **Columns:** base metadata (`source`, `relative_audio_path`, `speakerid`) plus `prompt1_speaker_idiosyncratic_style_instruction` through `prompt10_speaker_idiosyncratic_style_instruction`.
142
 
143
  ### VCTK speaker prompts
144
 
 
183
  | `ears` | [EARS Dataset](https://github.com/facebookresearch/ears_dataset) |
184
  | `expresso` | [Expresso](https://github.com/facebookresearch/textlesslib/tree/main/examples/expresso/dataset) |
185
  | `eval_giga` | [GigaSpeech](https://huggingface.co/datasets/speechcolab/gigaspeech) (evaluation subset) |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
data/In_domain_balanced_set/test-00000-of-00001.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7fa0e7f8e6c8949f3e96b36349522d9eaa9a82cb7ba38ee2a50e94b91b867ca3
3
- size 29880
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08703b00146ec58532a3f2d5cdeeffe89a4d9b96eac51d8ec9a5871d8c0e1a42
3
+ size 28049
data/StyleInstructCaps-MetaSet/test-00000-of-00001.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9242c2c4fbad451d2ea488b01a9035478bfedcad15a7615a85d892828bf8a0bf
3
- size 359180
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb4afc31fd35b743a07a35a028119f6f834d8573a15fe39d7cfcae5d04ae37a4
3
+ size 316893
data/StyleInstructCaps/eval-00000-of-00001.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:706394e1e721593dae5e930005f277298783e6be135176265677e29b61393c2c
3
- size 11134420
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:24c100471a38e1985df6e2ebec5cb21ec9414b426232520da3754683dbc8ff04
3
+ size 9961717
data/StyleInstructCaps/train-00000-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ed15af6ce2130ffb96df1cc78b60d802cee64ad877b06fdc6bbf4ece4969ce02
3
- size 70915714
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45f28bd4c47fb1e66a50ff785d5aeebf19355ae27e541764564a3c348028ca96
3
+ size 64629185
data/StyleInstructCaps/train-00001-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f462a1697cd83afab9d6a1550ce264102b8555585e0ed42a480b58f131939760
3
- size 70937964
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1d193ff3782f0f65fa580aefb8b99032fdc2de039eb8664b325dfc69ceaad768
3
+ size 64654933
data/StyleInstructCaps/train-00002-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:227b7890a7b29ebc3e0ebe7ff26dc46467d77b60aa242728d25c6e07a79a54b2
3
- size 71013201
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aeac2b6392cbdc8aa8738761713161bf04bc53f08cd79d0b651ecd5b45d2f5fe
3
+ size 64728994
data/StyleInstructCaps/train-00003-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:113cb0b459da5059594c770988a161e8139ab72538bccf882c1e7ee533a6b8a6
3
- size 70935589
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e88e773837860025a235f7ae42612582978f8c0e694d8b58ff5ce16fd6f07487
3
+ size 64649856
data/StyleInstructCaps/train-00004-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7039dc904bdf7a76a2487904f49d8d1f9e6c95ff6f9bca39770a171f83715e07
3
- size 71053586
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1eb0bd98d7eea5f09469efa876c674f4520d861da7586e295528bfe299cdd41d
3
+ size 64776387
data/StyleInstructCaps/train-00005-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7020c2990b678c43234bd97107331c75e061320d891b03bf476fee59d39f751c
3
- size 71001711
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba95fbafb1207541d05fb88d8f8d0869501b8ad6a071773f32062046278eaa5a
3
+ size 64724113
data/StyleInstructCaps/train-00006-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:42b436f9b2cc80c7da8f2fae0dec62ca1c8aeedd7af564dde400a7438c5d0415
3
- size 71012499
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fa94e9ec14fa3f69defb4f531ce7b1081fde86082952f6ab8f93bc142a08fb36
3
+ size 64738026
data/StyleInstructCaps/train-00007-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:19f3a9eec1189bc363cc1640af1b41a017e2a542d12bbc45fff9969e10986f33
3
- size 71038322
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:825b9883168e21084de28988f5a918d9ea9506cfedcaf850045be391cd41c6f2
3
+ size 64759883
data/StyleInstructCaps/train-00008-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ee8b61cc318dc9e5d35928cd4cd3d78abf076d2ac1c5b60b5ec10816a52d8a85
3
- size 70930255
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:01073126644e9591f84fc4ac9d81333fe671a76c75fea8c90d54d306bf7471ef
3
+ size 64664058
data/StyleInstructCaps/train-00009-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:23633c1dd6bbdeaf4de1fd702992772930761abe7dfa272acb869c24abcc4e5a
3
- size 72480944
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e67c3bd8a01382e2e443ad775a6a87417a001c97371292539e4f45dd07a2a11
3
+ size 64498536
data/StyleInstructCaps/train-00010-of-00011.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9591d310b66cc914e4c4bfd4e8865861032b2c8adb43f46162d5f769a658409f
3
- size 23162000
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb90e1df162c0733690c09a533508c481028e3284d71ea08de05b44b0e3b5086
3
+ size 20532521
scripts/prepare_release.py CHANGED
@@ -97,11 +97,9 @@ BASE_FIELDS = {
97
  "relative_audio_path",
98
  "transcription",
99
  "speakerid",
100
- "duration",
101
  }
102
 
103
  STYLE_CAPTION_FIELDS = {
104
- "text_description",
105
  "speaker_idiosyncratic_style",
106
  "situational_contextual_style",
107
  "expressive_emotional_style",
@@ -112,9 +110,7 @@ STYLE_CAPTION_FIELDS = {
112
  "captions",
113
  }
114
 
115
- STYLE_TEXT_FIELDS = STYLE_CAPTION_FIELDS - {"text_description"}
116
-
117
- LIST_FIELDS = {"text_description"}
118
 
119
 
120
  def _to_text(value: Any) -> str | None:
@@ -128,22 +124,6 @@ def _to_text(value: Any) -> str | None:
128
  return str(value)
129
 
130
 
131
- def _to_string_list(value: Any) -> list[str]:
132
- if value is None:
133
- return []
134
- if isinstance(value, str):
135
- return [value]
136
- if isinstance(value, list):
137
- return [str(item) for item in value if item is not None]
138
- return [str(value)]
139
-
140
-
141
- def _to_float(value: Any) -> float | None:
142
- if value is None:
143
- return None
144
- return float(value)
145
-
146
-
147
  def is_instruction_field(key: str) -> bool:
148
  return key.endswith("_instruction") or (
149
  key.startswith("prompt") and "_instruction" in key
@@ -163,13 +143,6 @@ def normalize_record(record: dict[str, Any]) -> dict[str, Any]:
163
  if field in cleaned:
164
  cleaned[field] = _to_text(cleaned[field])
165
 
166
- for field in LIST_FIELDS:
167
- if field in cleaned:
168
- cleaned[field] = _to_string_list(cleaned[field])
169
-
170
- if "duration" in cleaned:
171
- cleaned["duration"] = _to_float(cleaned["duration"])
172
-
173
  return cleaned
174
 
175
 
 
97
  "relative_audio_path",
98
  "transcription",
99
  "speakerid",
 
100
  }
101
 
102
  STYLE_CAPTION_FIELDS = {
 
103
  "speaker_idiosyncratic_style",
104
  "situational_contextual_style",
105
  "expressive_emotional_style",
 
110
  "captions",
111
  }
112
 
113
+ STYLE_TEXT_FIELDS = STYLE_CAPTION_FIELDS
 
 
114
 
115
 
116
  def _to_text(value: Any) -> str | None:
 
124
  return str(value)
125
 
126
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
127
  def is_instruction_field(key: str) -> bool:
128
  return key.endswith("_instruction") or (
129
  key.startswith("prompt") and "_instruction" in key
 
143
  if field in cleaned:
144
  cleaned[field] = _to_text(cleaned[field])
145
 
 
 
 
 
 
 
 
146
  return cleaned
147
 
148