---
pretty_name: YodaLingua English TTS Arena — Model Outputs & Human Preferences
license: other
language:
- en
task_categories:
- text-to-speech
size_categories:
- 1K<n<10K
tags:
- audio
- text-to-speech
- tts
- voice-cloning
- human-preferences
- pairwise-ranking
- benchmark
- english
- leaderboard
configs:
- config_name: outputs
  default: true
  data_files:
  - split: test
    path: data/outputs/*.parquet
- config_name: judgments
  data_files:
  - split: train
    path: data/judgments/*.parquet
- config_name: prompts
  data_files:
  - split: test
    path: data/prompts/*.parquet
- config_name: model_summary
  data_files:
  - split: test
    path: data/model_summary/*.parquet
---

# YodaLingua English TTS Arena

**1,080 generated clips · 18 TTS systems ·
301 anonymous pairwise judgments · 20 reference
speakers · 60 distinct prompts**

This is the public English snapshot of the
[YodaLingua multilingual TTS Arena](https://huggingface.co/spaces/Thomcles/yodalingua-tts-arena).
It pairs model outputs, shared zero-shot voice references, objective measurements, and
privacy-safe human preferences on the same frozen evaluation protocol.

## Start here

The default **`outputs`** configuration has one row per model × prompt, so the viewer
opens on **1,080 playable generated clips**, not a 60-row prompt table.

```python
from datasets import load_dataset

outputs = load_dataset(
    "Thomcles/YodaLingua-TTS-Arena-English",
    "outputs",
    split="test",
)
judgments = load_dataset(
    "Thomcles/YodaLingua-TTS-Arena-English",
    "judgments",
    split="train",
)
```

| Configuration | Rows | Intended use |
|---|---:|---|
| `outputs` | 1,080 | Listen, filter by model/task, recompute metrics |
| `judgments` | 301 | Pairwise ranking and preference research |
| `prompts` | 60 | Shared texts and reference voices |
| `model_summary` | 72 | Ready-to-use objective and early preference summaries |

## Evaluation design

- **Natural:** conversational rhythm, spontaneity, and varied sentence structure.
- **Expressive:** emotion, prosody, intention, and storytelling delivery.
- **Pronunciation:** proper nouns, rare words, loanwords, and difficult sequences.
- **Protocol:** 20 speakers × 3 unique prompts per speaker = 60 English trials per model.
- **Coverage:** 18 systems have complete English output coverage in this snapshot.
- **No cherry-picking:** model failures, repetitions, and hallucinations remain part of the data.

## What you can build

- Compare TTS and zero-shot voice-cloning systems on identical inputs.
- Audit pronunciation, naturalness, expressiveness, cloning similarity, and failure modes.
- Recompute WER, CER, speaker similarity, and ranking methods.
- Train or test preference models, evaluators, and leaderboard tooling.
- Create internal model-selection reports while retaining exact benchmark provenance.

## Configurations

### `outputs`

Contains the generated FLAC bytes, target text, anonymized speaker ID, task, immutable
model revision or provider alias, model-specific license, inference timing, WER/CER
components, ASR transcript, and speaker similarity. `rtf` is generation time divided by
audio duration. API timing includes network/provider latency and is not a controlled L4
comparison.

### `judgments`

Contains only the trial/task, randomized A/B model identities, `a`, `b`, or `tie`, and
the preferred model when applicable. It deliberately excludes session/viewer identifiers,
IP or device information, comparison IDs, playback events, and exact timestamps.

### `prompts`

Contains each reference FLAC and target text once. Speaker IDs are release-scoped,
deterministic pseudonyms; source video/channel identifiers are not published here.

### `model_summary`

Contains Overall, Natural, Expressive, and Pronunciation rows per model. Objective error
rates are corpus-weighted. Speaker similarity is the mean of successful trial scores.
Preference values use the same regularized Bradley–Terry method as the Arena.

## Human-preference limitations

The judgments are from self-selected anonymous listening sessions, not a representative
panel of verified English speakers. Sessions are not people. Early ratings can move
substantially as votes accumulate; consult judgment counts and uncertainty fields.
Pair scheduling is adaptive, so the raw data is not an independently and identically
distributed sample. Use this snapshot for exploration and transparent research, not as
proof of a universal “best” model.

## Licensing

The dataset-level license is **`other`** because generated outputs inherit different model
and provider terms. Every output and summary row includes `model_license`,
`model_repository`, and `model_revision`. **This dataset does not grant blanket commercial,
training, or redistribution rights for all audio.** Review the relevant model/provider
terms and the YodaLingua source dataset terms before reuse.

## Provenance

- Arena snapshot: `arena-v3-api-complete-2026-09-16`
- Voting cohort: `5cc0b0e569fdad5debdb78b8c36d5a6b7e387a82e28e2acfb6d1b13e96e8f326`
- Private source revision: `727ef53ca1c5b37ac43e02beeb6a8332ec3dbda8`
- Benchmark code: <https://github.com/Thomcle/yodalingua-tts-benchmark>
- Live Arena: <https://huggingface.co/spaces/Thomcles/yodalingua-tts-arena>
- Methodology: <https://github.com/Thomcle/yodalingua-tts-benchmark/blob/main/docs/task-methodology-v3.md>

The public export is deterministic for a fixed catalog, source revision, vote database
snapshot, exporter version, and PyArrow version. File checksums and row counts are recorded
in `release.json`.

## Citation

```bibtex
@misc{yodalingua_tts_arena_english,
  author = {YodaLingua contributors},
  title = {YodaLingua English TTS Arena: Model Outputs and Human Preferences},
  year = {2026},
  url = {https://huggingface.co/datasets/Thomcles/YodaLingua-TTS-Arena-English}
}
```