EmoTra-Ash-Transition-50K

Dataset summary

EmoTra-Ash-Transition-50K contains 50,000 synthetic English emotional-speech recordings for studying continuous emotion transitions in speech generation. Every recording uses the synthetic ash voice. A record starts from one EmoVoice-DB utterance A and is conditioned by a second EmoVoice-DB utterance B. The transcript is the text of A. Most source pairs use different emotion labels. A small number retain the same label because the deterministic VAD-pair sampler selects different utterances independently of the categorical label.

The dataset is designed for the EmoTra-TTS reproduction pipeline. It is suitable for research on emotion-conditioned speech synthesis, transition modeling, speech-token prediction, and analysis of synthetic emotional speech. It is not a benchmark of human emotional speech, speaker verification, or real-world voice conversion.

Property	Value
Records	50,000
Audio	24 kHz, mono WAV
Total duration	75.27 hours
Clip duration	1.36 to 12.80 seconds
Mean clip duration	5.42 seconds
Median clip duration	5.40 seconds
Voice	ash, a synthetic EmoVoice-DB speaker timbre
Text language	English
Source-pair emotion types	41
Mean Whisper character error rate	0.0073
Maximum accepted Whisper character error rate	0.1000

Data sources

The source corpus is yhaha/EmoVoice-DB at revision e0e61055fc5c2b52bc20d8b45f8ce3037228a61d. EmoVoice-DB is labelled MIT. Its dataset card describes GPT-4o-generated text and emotion descriptions, GPT-4o-audio-generated speech, and five synthetic speaker timbres. This release uses only the ash timbre.

Transition audio was generated with FunAudioLLM/CosyVoice2-0.5B, labelled Apache-2.0. The release contains generated audio and metadata. It does not include EmoVoice-DB source WAV files, CosyVoice model weights, or source-model checkpoints.

Construction

1. Filter the EmoVoice-DB ash training entries using the reproduction pipeline's VAD selection.
2. Form deterministic source-pair candidates from distinct source utterances whose VAD trajectories meet the reproduction pipeline's change criterion.
3. Generate speech tokens from A's text and prompt audio with CosyVoice2.
4. Synthesize mel trajectories conditioned on A and B speaker embeddings.
5. Decode the mel trajectories with the CosyVoice HiFT vocoder.
6. Transcribe generated WAVs with Whisper.
7. Retain a recording only when its character error rate against A's text is at most 0.10.
8. Merge terminal outcomes from recovery and production shards. When both contain the same source-pair index, retain the newer production-shard outcome. Sort accepted indices in ascending order and retain the first 50,000.

The transition direction is A to B. The audio does not contain a concatenation of the source recordings. It is a generated waveform conditioned by the paired source information.

Quality checks

Every released row passed these checks before publication:

• one accepted terminal synthesis outcome;
• a unique source-pair index;
• an existing WAV file;
• 24 kHz sample rate and one audio channel;
• successful WAV decode with finite samples;
• populated transcript, source keys, VAD metadata, and ASR result;
• Whisper character error rate at most 0.10.

The checks verify file integrity and text agreement. They do not establish human-rated naturalness, emotional correctness, safety, factual accuracy, or performance on any downstream task.

Data layout

.
├── README.md
├── transition_data.jsonl
├── synthesis_manifest.json
├── dataset_statistics.json
├── train/
│   ├── metadata.jsonl
│   ├── transition_data.jsonl
│   └── audio/<index-range>/*.wav
└── held_out/
    ├── metadata.jsonl
    ├── transition_data.jsonl
    └── audio/<index-range>/*.wav

The Dataset Viewer exposes two splits. train contains 45,000 records. held_out contains 5,000 records. Each split's metadata.jsonl stores an audio column declared as an audio feature and a matching audio/<index-range>/<name>.wav file_name, relative to its split directory. The Hub renders the audio column as a playable preview. The index-range directories each contain at most 10,000 WAV files, which satisfies the Hub repository directory limit.

The split uses a fixed seed of 0. The 50,000 records are first grouped by the ordered pair (source_a_key, source_b_key). The pair groups are shuffled once. Whole groups are assigned to held_out until it contains 5,000 records. The remaining 45,000 records form train. A transition record and an identical ordered A/B source pair never appear in both splits.

This split is record-disjoint and source-pair-disjoint, not source-utterance-disjoint. The 50,000 transition records reuse 1,067 source EmoVoice-DB utterances across different pairs. A source utterance may therefore condition one train record and one held-out record. The held-out split measures generalization to unseen transition pairs. It does not measure generalization to unseen source utterances or unseen speakers.

Root transition_data.jsonl contains the 50,000 full records in source-pair-index order with split-relative audio paths. The split-local transition_data.jsonl files contain the corresponding 45,000 and 5,000 records. synthesis_manifest.json is the terminal accepted-outcome manifest keyed by source-pair index. dataset_statistics.json records the construction-time summary values used in this card.

metadata.jsonl fields

Field	Description
audio	Playable audio feature in the Hub Dataset Viewer
file_name	Split-relative WAV path that resolves the audio feature
key	Stable transition key, transition_<source_pair_index>
text	Source A transcript and expected generated content
speaker	Synthetic speaker timbre, always ash in this release
source_pair_index	Deterministic global pair index used for selection
source_a_*	Source A key, emotion, and VAD values
source_b_*	Source B key, emotion, and VAD values
asr_cer	Whisper character error rate against text
asr_transcription	Whisper transcription of the generated recording

Emotion-pair counts

Source A emotion	Source B emotion	Records
angry	disgusted	1,133
angry	fearful	858
angry	happy	3,429
angry	neutral	1,994
angry	sad	5,550
angry	surprised	656
disgusted	angry	1,198
disgusted	happy	2,030
disgusted	neutral	129
disgusted	sad	224
disgusted	surprised	197
fearful	angry	827
fearful	fearful	1
fearful	happy	666
fearful	neutral	27
fearful	sad	7
fearful	surprised	42
happy	angry	3,618
happy	disgusted	2,134
happy	fearful	719
happy	happy	32
happy	neutral	419
happy	sad	4,781
happy	surprised	242
neutral	angry	1,980
neutral	disgusted	124
neutral	fearful	31
neutral	happy	422
neutral	sad	202
sad	angry	5,568
sad	disgusted	196
sad	fearful	3
sad	happy	4,857
sad	neutral	185
sad	surprised	2,137
surprised	angry	670
surprised	disgusted	167
surprised	fearful	52
surprised	happy	270
surprised	sad	2,219
surprised	surprised	4

Intended use

• Research on synthetic emotional speech generation.
• Training and evaluation of models that use text, audio, VAD values, or emotion-transition metadata.
• Reproduction and analysis of EmoTra-TTS-style transition generation.

Out-of-scope use

• Representing real people, real speaker identity, or authentic human emotional behavior.
• Biometric identification, authentication, surveillance, or speaker-verification claims.
• Clinical, psychological, hiring, education, or other high-impact decisions about people.
• Claiming that the generated voice is a recording of a person or that its emotional state is human ground truth.

Limitations and safety

All released speech is synthetic. The ash label identifies a synthetic timbre from EmoVoice-DB, not a real person. Generated speech may contain intelligibility errors, unnatural prosody, artifacts, unintended emotional cues, or text not faithfully reflected in the waveform. The Whisper CER filter measures transcription agreement only. It does not measure emotional transition quality.

The source text and audio were generated by upstream systems. Users should preserve this synthetic-data context in downstream documentation and should not present the recordings as human speech. Check the laws, platform terms, and policies applicable to the intended use and distribution location.

License and attribution

This dataset is released under the MIT License. Preserve the source attribution when redistributing or creating derivatives:

• Source data: EmoVoice-DB, MIT, revision e0e61055fc5c2b52bc20d8b45f8ce3037228a61d.
• Transition generator: CosyVoice2-0.5B, Apache-2.0.

The source dataset, generated dataset, and model are separate distributions. This repository does not redistribute source audio, model weights, or source-model checkpoints.

Citation

If you use this dataset, cite EmoVoice and CosyVoice as appropriate for the source data and generator:

@article{yang2025emovoice,
  title={EmoVoice: Pushing the Boundaries of Open-Source Emotional Text-to-Speech},
  author={Yang, Haoran and others},
  journal={arXiv preprint arXiv:2504.12867},
  year={2025}
}

@article{du2024cosyvoice,
  title={CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer Based on Supervised Semantic Tokens},
  author={Du, Zhihao and others},
  journal={arXiv preprint arXiv:2407.05407},
  year={2024}
}