README.md

---
license: cc-by-4.0
task_categories:
  - audio-classification
  - text-to-speech
language:
  - en
pretty_name: "E-VOC: Expressive VOice Control Corpus"
tags:
  - text-to-speech
  - instruction-guided-tts
  - expressive-speech
  - human-perception
  - human-annotation
  - emotion
  - prosody
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: task1_adv_degree
        path: data/task1_adv_degree/**
      - split: task2_emotion_intensity
        path: data/task2_emotion_intensity/**
      - split: task3_emphasis
        path: data/task3_emphasis/**
      - split: task4_age
        path: data/task4_age/**
      - split: task5_emotion
        path: data/task5_emotion/**
---

# E-VOC: Expressive VOice Control Corpus

**E-VOC** (Expressive VOice Control) is a large-scale human-evaluation corpus for
studying the **instruction-perception gap** in instruction-guided text-to-speech
(ITTS) systems. It pairs synthesized speech from five ITTS systems with
large-scale human ratings across several expressive dimensions, so that the
alignment between a user's style instruction and what listeners actually perceive
can be measured.

This corpus accompanies the paper:

> **Do You Hear What I Mean? Quantifying the Instruction-Perception Gap in
> Instruction-Guided Expressive Text-To-Speech Systems**
> Yi-Cheng Lin, Huang-Cheng Chou, Tzu-Chieh Wei, Kuan-Yu Chen, Hung-yi Lee.
> Accepted to **ICASSP 2026**. arXiv:[2509.13989](https://arxiv.org/abs/2509.13989).

## Overview

ITTS lets users control speech generation through natural-language prompts, but
how well listeners perceive the requested style is largely unexplored. E-VOC
provides a perceptual analysis of ITTS controllability across two expressive
dimensions (**adverbs of degree** and **graded emotion intensity**) and adds
human ratings on **speaker age** and **word-level emphasis**.

## ITTS systems

Audio is generated by five systems and stored under the corresponding
top-level directories in this repository:

- `Parler-TTS-large-v1`
- `Parler-TTS-mini-v1`
- `PromptTTS++`
- `UniAudio`
- `gpt-4o-mini-tts`

## Annotation tasks

The dataset is exposed as **five splits** (one per task) of the `default` config.
Select a split in the Dataset Viewer to browse that task; each row pairs the
playable `audio` clip with its annotation.

| Viewer split | Dimension | What annotators judged |
| --- | --- | --- |
| `task1_adv_degree` | Adverbs of degree | Perceived intensity (Very Low … Very High) of an adverb-modulated emotion clip |
| `task2_emotion_intensity` | Graded emotion intensity | Perceived intensity of an emotion within an emotion sub-category |
| `task3_emphasis` | Word-level emphasis | Which word in the sentence sounds emphasized |
| `task4_age` | Speaker age | Perceived speaker age |
| `task5_emotion` | Emotion classification | Perceived emotion category of the clip |

Tasks 1 and 5 are annotated on the **same** `Adv/emotion` audio clips, asking a
different question per task (degree vs. emotion category).

Each split is an [AudioFolder](https://huggingface.co/docs/datasets/audio_dataset#audiofolder)
under `data/<split>/` (a `metadata.csv` plus the split's clips), so the viewer
shows an inline audio player and duration distribution next to every annotation.
The same content is also available as plain CSVs under
[`labels/`](./labels) for easy download.

## Dataset statistics

| File | Annotations | Unique clips | Annotators | Annotation values |
| --- | --- | --- | --- | --- |
| `Task1_Adv_Degree.csv` | 17,482 | 2,880 | 29 | `1 - Very Low` … `5 - Very High`, `Unclear` |
| `Task2_Emotion_Intensity.csv` | 29,295 | 3,600 | 59 | `1 - Very Low` … `5 - Very High`, `Unclear` |
| `Task3_Emphasis.csv` | 10,811 | 1,440 | 27 | emphasized word (50 distinct) |
| `Task4_Age.csv` | 3,597 | 720 | 10 | `Child`, `Teenager`, `Adult`, `Elderly`, `Unclear` |
| `Task5_Emotion.csv` | 20,205 | 2,880 | 40 | `Angry`, `Happy`, `Sad`, `Surprised`, `Neutral`, `Other`, `Unclear` |
| **Total** | **81,390** | — | **144 unique** | — |

Each clip is rated by multiple annotators. Generation metadata spans 5 ITTS
systems x 3 samples x 2 templates x several conversational contexts
(e.g. Customer, Family, Friends, Lover, Teacher-Student, Normal).

## Columns

Each row (in both the viewer splits and the CSVs) has:

- `audio` – the playable audio clip.
- `Task` – task identifier.
- `Model`, `Context`, `Template`, `Sample` – generation metadata.
- `Ground Truth` – the intended/target attribute of the clip. Its meaning is
  task-specific: emotion + degree for Task 1 (e.g. `Slightly Sad`), emotion
  intensity for Task 2 (e.g. `2 - Low`), the emphasized word for Task 3, the
  target age for Task 4, and the target emotion for Task 5.
- `Annotation` – the annotator's response.
- `Annotator ID` – anonymized annotator pseudonym.
- `FileName` – the original synthetic clip identifier used during collection.
- `Sentence` – the carrier sentence read in the clip (present in all tasks).

In the CSV files the audio link is the `audio_path` column, a repo-relative path
pointing into `data/<split>/`.

## Repository layout

- `data/<split>/` – the human-annotated clips (one AudioFolder per task) plus the
  per-split `metadata.jsonl`. This is the full audio backing every annotation.
- `acoustic/<model>/…` – the objective acoustic stimuli for Task I (Adverbs of
  Degree): `Adv/pitch`, `Adv/loudness`, and `Adv/rate` clips (5,400 total). These
  are analyzed objectively (LUFS / F0 / words-per-second, paper Fig. 1) and have
  no human ratings.
- `labels/` – the same annotations as plain CSVs.

## Usage

```python
from datasets import load_dataset

# Load a single task split, with decoded audio
ds = load_dataset("wizzzzzzzzz/E-VOC", split="task1_adv_degree")
row = ds[0]
print(row["Annotation"], row["Sentence"])
print(row["audio"]["sampling_rate"], row["audio"]["array"].shape)

# Or load all five task splits at once
all_tasks = load_dataset("wizzzzzzzzz/E-VOC")
print(all_tasks)  # task1_adv_degree, task2_emotion_intensity, ... task5_emotion
```

## Privacy / anonymization

Original annotator identifiers (Prolific and Amazon Mechanical Turk IDs) are
**not** published. Each `Annotator ID` is a salted `HMAC-SHA256` pseudonym
(`anon_` + 12 hex chars). The same annotator maps to the same pseudonym across
all five files, enabling cross-task analysis, while the secret salt is kept
private so the pseudonyms cannot be reversed to real identifiers.

## Citation

```bibtex
@inproceedings{lin2026you,
  title={Do You Hear What I Mean? Quantifying the Instruction-Perception GAP in Instruction-Guided Expressive Text-to-Speech Systems},
  author={Lin, Yi-Cheng and Chou, Huang-Cheng and Wei, Tzu-Chieh and Chen, Kuan-Yu and Lee, Hung-yi},
  booktitle={ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={16472--16476},
  year={2026},
  organization={IEEE}
}
```