Datasets:
The dataset viewer is not available for this split.
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Decider Audio – VAST (spoken)
Part of the Decider Audio collection.
The model hears an opinion post, reads a topic as text, and must decide whether the post is against the topic, in favor of it, or neutral. There are three options. The topic is text and is often not literally in the post, so the words of the post alone cannot give the answer.
It has 18,545 rows from VAST, the zero-shot stance detection dataset. 3,313 distinct posts were rendered by 6 text-to-speech (TTS) models, about 33.24 hours of audio. The target is answer, an option key A, B or C. answer_label is against, favor or neutral.
Suggested prompt (the model sees the audio at the marked position):
Post:
<audio>
Topic: public schools
Determine the stance of the post toward the topic.
A: against
B: favor
C: neutral
Answer with the option key.
spoken_text is the original post, included for quality control and ASR baselines. Do not give it to a model as input when training an audio model.
| Split | Rows | Spoken clips | Distinct voices | Hours | Re-rendered | Failing checks | Answers |
|---|---|---|---|---|---|---|---|
train |
13,477 | 1,845 | 204 | 18.58 | 8 | 0 | 5,595 against, 5,327 favor, 2,555 neutral |
validation |
2,062 | 682 | 195 | 6.68 | 2 | 0 | 697 neutral, 684 against, 681 favor |
test |
3,006 | 786 | 193 | 7.98 | 1 | 0 | 1,044 neutral, 1,018 against, 944 favor |
Answers are counted per row. No speaker appears in more than one split. test must not be used for training. Most VAST topics in validation and test do not occur in train (zero-shot setting, see seen).
How it is made
- Text.
data/VAST/vast_{train,dev,test}.csvofemilyallaway/zero-shot-stanceat a pinned commit (SHA-256 verified). It has 13,477, 2,062 and 3,006 rows (devis published asvalidation). Instruction, options and keys follow the Decision Index projection of VAST. - One clip per post. VAST asks for the stance of the same post toward several topics. Rows of a split with the same post share one clip (
spoken_unit,spoken_unit_rows> 1), so the audio is identical and only the topic differs. 3,102 clips serve more than one row, and 18,334 rows share their clip. - Content-only speech. Only the post is spoken. Topic, instruction, options and answer never enter the audio. Posts are long user comments, and 2,351 clips above 400 characters are spoken in segments. 4 clips got a cleaned
tts_text(lists and ID tokens without brackets and quotes). All others got the post verbatim. - Model and voice assignment. Six TTS models each render the same number of clips (±1) per split, and the assignment does not look at the answer. Texts are classified (
speech_category) as in the When2Call dataset of this collection. Texts with lists, URLs, ID tokens, handles or long numbers go preferably to the model that spoke that type best in the When2Call pilot test (lists to Sopro, URLs and handles to Supertonic, ID tokens to Qwen, long numbers to OmniVoice, then Pocket). Plain texts fill the rest. The voice is picked by a seeded hash from the pool of the matching split.
| Question type | Spoken as | Ranking from the pilot | Clips | chatterbox |
omni |
pocket |
qwen |
sopro |
supertonic |
|---|---|---|---|---|---|---|---|---|---|
| handle | original | supertonic > omni > qwen > sopro > chatterbox > pocket | 5 | 0 | 0 | 0 | 0 | 0 | 5 |
| url | original | supertonic > omni > chatterbox > qwen > sopro > pocket | 5 | 0 | 0 | 0 | 0 | 0 | 5 |
| list | cleaned | sopro > pocket > chatterbox > qwen > supertonic > omni | 2 | 0 | 0 | 0 | 0 | 2 | 0 |
| id token | cleaned | qwen > omni > pocket > chatterbox > sopro > supertonic | 7 | 0 | 0 | 0 | 7 | 0 | 0 |
| long number | original | omni > pocket > qwen > supertonic > chatterbox > sopro | 346 | 0 | 346 | 0 | 0 | 0 | 0 |
| plain | original | balanced | 2,948 | 552 | 206 | 552 | 544 | 551 | 543 |
- Voices. Voice-cloning models use English reference clips from a private voice collection (not redistributed). Speakers are assigned to splits before any voice is chosen, so test speakers never occur in training. Preset-voice models use the presets of their split's partition. There are 592 distinct voices in total.
- Audio. 24 kHz mono, 16-bit FLAC. Clips whose peak exceeded full scale were scaled to 0.99 (
audio_gain). Flat-topped (clipped) renders were rejected instead. - Re-rendering, never removal. Every render was checked. It fails if the ASR word error rate is above 0.5, if it has more than 30 characters per second, or if it has 3 or more extra words that the text does not contain (10 % of its words for texts longer than 30 words). A failed clip was rendered again with another voice and seed on the same model (up to 5 attempts). 11 clips carry a re-render. 0 clips (0 rows) still failed after the last attempt and keep their best render with
passes_checks = false. A second repair pass rendered 124 clips again. 18 of them still failed a check and 106 passed the word check but leave out part of their text. A clip got up to 10 more attempts on its own model. Clips on Supertonic and clips that a model could not render at all then moved to another model for up to 3 more. A clip now also passes when its only problem is the word error rate or extra words and its transcript sounds like the text. This check compares phonemes after numbers and units are spelled out, so "topup" and "top up" or "1960" and "nineteen sixty" match. A clip that leaves out part of its text does not pass this way. A truncated clip is replaced only when the new render leaves out less.
TTS models
| TTS model | Checkpoint | Voice mode | Clips | Distinct voices | Hours | Mean WER | Exact | WER ≤ 0.2 | Re-rendered |
|---|---|---|---|---|---|---|---|---|---|
chatterbox |
ResembleAI/chatterbox | clone | 552 | 349 | 5.04 | 0.0244 | 26.6% | 99.6% | 7 |
omni |
k2-fsa/OmniVoice | clone | 552 | 332 | 6.72 | 0.0209 | 27.5% | 100.0% | 1 |
pocket |
kyutai/pocket-tts | clone | 552 | 342 | 5.26 | 0.0184 | 34.8% | 99.8% | 0 |
qwen |
Qwen/Qwen3-TTS-12Hz-0.6B-Base | clone | 551 | 336 | 5.05 | 0.0225 | 28.5% | 100.0% | 2 |
sopro |
samuel-vitorino/sopro-v2-turbo | clone | 553 | 354 | 5.67 | 0.0206 | 33.1% | 99.8% | 1 |
supertonic |
Supertone/supertonic-3 | preset | 553 | 10 | 5.50 | 0.0272 | 25.1% | 100.0% | 0 |
Counts are spoken clips.
Checks
Every clip was transcribed with Systran/faster-whisper-base.en and compared with spoken_text after lower-casing and removing punctuation. Numbers are compared digit by digit after number words become digits ("twenty-five" = "25", "$3" = "3 dollars", "50%" = "50 percent").
- The mean WER over clips is 0.0223. 970 clips (29.3%) transcribe exactly. The 90th percentile is 0.055 and the 99th is 0.115.
- 99.9% of the clips have WER ≤ 0.2. Filter on
asr_werorpasses_checksfor a cleaner subset. - Agreement with the transcript shows that the words were spoken. It does not show that the voice sounds natural or resembles the reference voice. Only a sample of 10 texts from the pilot was reviewed by listening.
Limitations:
- Synthetic speech only. There is no human speech. Expect a domain gap to real speakers, accents, noise and telephone channels.
- Long clips. Posts are long written comments. Heard at once, they carry more detail than a typical spoken turn.
- Topics stay text. The model must relate a written topic to a spoken post.
- English only. Speaker attributes are not annotated.
Fields
| Field | Type | Description |
|---|---|---|
id |
string | vast-<split>-<index>, where index is the row of the source CSV |
audio |
Audio, 24 kHz mono FLAC | Spoken post |
instruction |
string | Topic: <topic> and the task text |
choices, choice_labels |
list of strings | against, favor, neutral |
choice_keys |
list of strings | A, B, C |
answer |
string | Option key of the stance |
answer_index |
int | 0 = against, 1 = favor, 2 = neutral (VAST's label) |
answer_label |
string | The stance |
spoken_text |
string | The post as in the source |
spoken_language |
string | Always en |
tts_text |
string | Text the TTS model got (see "How it is made") |
speech_category |
string | Text type that chose the model: list, url, id_token, handle, long_number or plain |
source_dataset, source_id, source_split |
string | Provenance |
source_metadata |
JSON string | Further source fields (see below) |
spoken_unit, spoken_unit_rows |
string, int | Id of the spoken clip and the number of rows that share it |
tts_model |
string | Project key of the TTS model |
tts_repo, tts_revision, tts_sdk_revision |
string | Checkpoint, pinned revision and runtime revision |
voice_id |
string | Reference-voice id (clone models) or preset name |
voice_kind |
clone or preset |
How the voice was chosen |
voice_split |
string | Split the voice belongs to, always equal to the row's split |
tts_settings |
JSON string | Generation settings |
tts_seed |
int | Seed of the generation call |
duration_seconds, n_segments |
float, int | Length of the clip and number of segments. Texts above 400 characters are split at whitespace into ordered segments joined with 180 ms of silence (2,351 clips) |
audio_gain |
float | 1.0 unless the clip overshot full scale and was scaled down (507 clips) |
asr_text, asr_wer, asr_metric, asr_model |
string, float, string, string | Round-trip speech recognition check, see "Checks" |
render_attempt |
int | 0 for the first render, higher if the text was rendered again after a failed check |
passes_checks |
bool | False if even the last attempt failed a check |
source_metadata holds VAST's topic, new_id, ori_id, ori_topic, type_idx, seen and contains_topic. Where the source file has them, it also holds the challenge annotations Qte, Sarc, Imp, mlS and mlT.
Clips are 0.54–128.31 s long (mean 36.12 s per clip).
License and attribution
- Text. VAST (Allaway and McKeown, 2020), from the authors' repository
emilyallaway/zero-shot-stance. The repository states no license at the pinned commit. The posts are comments from The New York Times "Room for Debate" section, collected in the argument reasoning comprehension corpus (Habernal et al., 2018). Check these terms before any use beyond private research. - Speech. OmniVoice weights are reported as CC BY-NC 4.0 (verify on
k2-fsa/OmniVoice). Qwen3-TTS is Apache-2.0, Sopro is Apache-2.0, Supertonic is OpenRAIL (use restrictions), Pocket TTS is CC BY 4.0 (attribution, and Kyutai's PyTorch weights are gated behind their terms) and Chatterbox is MIT.tts_reposhows which checkpoint rendered a clip. - Reference voices. Private collection, not redistributed.
The dataset is marked license: other. Do not use it commercially before these terms are resolved.
@inproceedings{allaway-mckeown-2020-zero,
title = {Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations},
author = {Allaway, Emily and McKeown, Kathleen},
booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
year = {2020},
url = {https://aclanthology.org/2020.emnlp-main.717}
}
- Downloads last month
- 24