README.md

metadata

dataset_info:
  - config_name: tts
    features:
      - name: id
        dtype: string
      - name: text
        dtype: string
      - name: prompt_text
        dtype: string
      - name: prompt_audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: speaker_id
        dtype: string
      - name: source_id
        dtype: int64
      - name: prompt_id
        dtype: int64
      - name: source_split
        dtype: string
    splits:
      - name: fa
        num_bytes: 336194003
        num_examples: 2098
    download_size: 6871992
    dataset_size: 336194003
  - config_name: tts_validation
    features:
      - name: id
        dtype: string
      - name: text
        dtype: string
      - name: prompt_text
        dtype: string
      - name: prompt_audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: speaker_id
        dtype: string
      - name: source_id
        dtype: int64
      - name: prompt_id
        dtype: int64
      - name: source_split
        dtype: string
    splits:
      - name: fa
        num_bytes: 296643213
        num_examples: 1851
    download_size: 6209505
    dataset_size: 296643213
configs:
  - config_name: tts
    data_files:
      - split: fa
        path: tts/fa-*
  - config_name: tts_validation
    data_files:
      - split: fa
        path: tts_validation/fa-*
language:
  - fa
source_datasets:
  - shekar-ai/Neyshekar
task_categories:
  - text-to-speech
license: cc0-1.0

Neyshekar Persian TTS evaluation

Reformatted from shekar-ai/Neyshekar, revision 30696ada746a66bb3047c984744b5ecfd60fe4e9, under CC0 1.0. Source authors: Ahmad Amirivojdan, Farzad Nadiri, Abolfazl Alizadeh and Shaghayegh Yaraghi. See Neyshekar (2026). Do not attempt to identify speakers.

When requested, tts/fa uses the predefined test split and tts_validation/fa uses validation, for development only. No training recordings are used. Requested source splits are checked for disjoint speaker IDs. This does not establish disjointness from any evaluated model's training data; model authors must disclose Neyshekar training overlap.

Within each split, select one 3–10 second prompt per speaker, closest to 5 seconds, ties broken by ascending source ID. Targets are the remaining 1–30 second recordings' transcripts, deduplicated per speaker. All selected prompt IDs and normalized prompt texts are excluded from targets across that split. Speakers without a prompt and a distinct target are omitted. Pairing uses no ASR/TTS quality scores. Rows are ordered by speaker ID, then numeric target ID. Speaker distribution follows the source and is not balanced. This is an automatically constructed evaluation set, not a manually audited TTS benchmark.

Changes: prompt/target pairing, filtering, and conversion to the leaderboard's id, text, prompt_text, prompt_audio schema, with source/prompt IDs and split provenance. Original transcripts and prompt WAV bytes are unchanged. Target recordings are not copied into the evaluation set. Preparation script: prepare_neyshekar_eval.py.

Scoring: Qwen3-ASR (fa), Persian normalized WER, RTFx, and WavLM SIM for cloning. UTMOS22 is English-only and is not reported here. Normalization is applied only at scoring time, never to the synthesis input. See the repository's Persian evaluation docs.

Preparation report

The SHA-256 covers ordered [target_id, prompt_id] pairs, encoded as compact JSON.

{
  "tts": {
    "source_split": "test",
    "source_rows": 2149,
    "prompt_id_or_text_excluded": 31,
    "duplicate_target_text": 20,
    "output_rows": 2098,
    "output_speakers": 30,
    "pairs_sha256": "b170ed14adc3718ed9fb2d958456c203c3127c0a528e4d0ed7b120bd49414bfe"
  },
  "tts_validation": {
    "source_split": "validation",
    "source_rows": 1886,
    "prompt_id_or_text_excluded": 28,
    "duplicate_target_text": 7,
    "output_rows": 1851,
    "output_speakers": 26,
    "pairs_sha256": "24ce4e740a466a478421bd90a1bbabf8627fb8e12d1ee1cfa55937a49796c757"
  }
}

Citation

All credit for the recordings and transcripts goes to the Neyshekar authors. If you use this evaluation set, please cite the source corpus:

@article{amirivojdan2026neyshekar,
  title={Neyshekar: An Open Persian Read-Speech Corpus for Automatic Speech Recognition},
  author={Amirivojdan, Ahmad and Nadiri, Farzad and Alizadeh, Abolfazl and Yaraghi, Shaghayegh},
  journal={arXiv preprint arXiv:2609.14542},
  year={2026},
  doi={10.48550/arXiv.2609.14542},
  url={https://arxiv.org/abs/2609.14542}
}