| --- |
| license: cc-by-4.0 |
| language: [en] |
| pretty_name: "EmoSpoof-TTS" |
| task_categories: [audio-classification] |
| size_categories: [10K<n<100K] |
| configs: |
| - config_name: default |
| data_files: |
| - {split: test, path: "data/test-*.parquet"} |
| tags: |
| - anti-spoofing |
| - audio-deepfake-detection |
| - speech |
| - benchmark |
| - arena-ready |
| paperswithcode_id: |
| arxiv: ["2505.23962"] |
| --- |
| |
| # EmoSpoofTTS |
|
|
| A **spoof-only** attack corpus of emotional text-to-speech (TTS) synthesis: |
| 36,000 clips spanning 3 modern TTS systems, 10 speakers, and 4 emotions, all |
| synthesized from transcripts of the Emotional Speech Dataset (ESD). |
|
|
| ## Overview |
|
|
| EmoSpoof-TTS (Mahapatra et al., "Can Emotion Fool Anti-spoofing?", Interspeech |
| 2025, arXiv:2505.23962) was built to study whether emotionally expressive TTS |
| is harder for anti-spoofing systems to detect than neutral TTS. For 10 ESD |
| speakers (`0011`-`0020`) and 4 emotions (Neutral, Angry, Sad, Happy), 300 |
| held-out ESD transcripts per speaker/emotion were synthesized with three |
| zero-shot TTS systems: |
|
|
| - **StyleTTS2** (Li et al., NeurIPS 2023) |
| - **F5-TTS** (Chen et al., arXiv:2410.06885) |
| - **CosyVoice** (Du et al., arXiv:2407.05407) |
|
|
| This gives 3 models x 10 speakers x 4 emotions x 300 utterances = 36,000 |
| synthetic clips (~29 hours), all parallel across models (same transcript per |
| `AUDIO_ID`, differing only by a `t{1,2,3}` model suffix). |
|
|
| ### Spoof-only caveat |
|
|
| **This dataset contains no bonafide audio** (`n_bonafide = 0`). The |
| corresponding ESD bonafide recordings are a separate, independently licensed |
| corpus and are not redistributed here. As a result, EER computed against this |
| dataset alone is degenerate/one-sided (all trials are target=spoof), so the |
| Arena scores it with the **1-SRR metric (`srr_complement`, lower is better)** |
| instead — the fraction of these emotional-TTS spoofs **not rejected** at a |
| fixed operating threshold `t*` transferred from each model's own DeepVoice EER |
| operating point (carried in the submission's `calibration` block). It is an |
| **attack-difficulty probe**: it measures how often a detector calibrated |
| elsewhere mis-classifies these emotional-TTS spoofs as bonafide (its miss rate |
| on this specific attack family). |
| |
| ## License & redistribution |
| |
| Released by the source authors under **Creative Commons Attribution 4.0 |
| International (CC BY 4.0)**, per `/mnt/datasets/emospoof/README.txt` section |
| 5 ("Copying"). The verbatim license text is in `LICENSE.txt`. |
| |
| ## Schema |
| |
| | Field | Type | Description | |
| |---|---|---| |
| | path | string | `{model}/{speaker}/{emotion}/<filename>.wav`, e.g. `StyleTTS2/0011/Neutral/0011_000001t1.wav`. | |
| | audio | Audio(16kHz mono) | Resampled on decode. | |
| | label | ClassLabel[bonafide, spoof] | Always `spoof` (index 1) — see spoof-only caveat above. | |
| | notes | string (JSON) | `{"utterance_id", "speaker_id", "emotion", "tts_model", "transcript"}`. | |
| |
| ## Quick Start |
| |
| ```python |
| from datasets import load_dataset |
| ds = load_dataset("SpeechAntiSpoofingBenchmarks/EmoSpoofTTS", split="test") |
| ``` |
| |
| ## Stats |
| |
| | n_total | n_bonafide | n_spoof | total duration | |
| |---|---|---|---| |
| | 36000 | 0 | 36000 | ~29h 51m | |
| |
| ## Source provenance |
| |
| - Source: `/mnt/datasets/emospoof/{StyleTTS2,F5TTS,CosyVoice}/<speaker>/<emotion>/*.wav` |
| from the EmoSpoof-TTS v1.0 release (Mahapatra et al., 2025, JHU SMILE Lab). |
| - Transcript and emotion metadata recovered from |
| `/mnt/datasets/emospoof/wav_list/<speaker>.txt` |
| (`AUDIO_ID\tTRANSCRIPT\tEMOTION`), matched to each clip by stripping the |
| trailing model suffix (`t1`=StyleTTS2, `t2`=F5-TTS, `t3`=CosyVoice) from the |
| filename stem. |
| - All 36,000 clips are `label=spoof`; no bonafide audio is included (see |
| spoof-only caveat above). |
| |
| ## Evaluation |
| |
| See `eval.yaml` and `submissions/README.md`. This dataset is spoof-only and is |
| scored with `srr_complement` (1-SRR, lower is better) via a DeepVoice threshold |
| transfer; the legacy `eer_percent` is degenerate on this dataset alone and is |
| interpreted as a miss-rate / attack-difficulty probe rather than a |
| conventional EER. |
| |
| ## Citation |
| |
| **Original paper**: Mahapatra, A., Ulgen, I.R., Naini, A.R., Busso, C., Sisman, |
| B. "Can Emotion Fool Anti-spoofing?" Interspeech 2025. arXiv:2505.23962. |
| |
| ```bibtex |
| @article{mahapatra2025emotion, |
| title = {Can Emotion Fool Anti-spoofing?}, |
| author = {Mahapatra, Aurosweta and Ulgen, Ismail Rasim and Naini, Abinay Reddy and Busso, Carlos and Sisman, Berrak}, |
| journal = {arXiv preprint arXiv:2505.23962}, |
| year = {2025} |
| } |
| ``` |
| |
| ## Maintainer |
| |
| Maintained by Kirill Borodin (SpeechAntiSpoofingBenchmarks). |
| - Email: ~~k.n.borodin@mtuci.ru~~ (deprecated — use kborodin.research@gmail.com) |
| - Telegram: [@korallll_ai](https://t.me/korallll_ai) |
| |
