--- license: cc-by-4.0 task_categories: - text-to-speech - reinforcement-learning language: - en tags: - audio - human-feedback - human-preferences - preference-learning - pairwise-comparison - reward-model - rlhf - dpo - text-to-speech - speech-synthesis - arena - leaderboard - elo - benchmark size_categories: - 100K- You agree to use this dataset under the CC-BY-4.0 license with attribution to Datapoint AI, and you agree not to attempt to re-identify annotators. Usage rights for the audio are governed by each model provider's terms. extra_gated_fields: First Name: text Last Name: text Email: text Affiliation (company or institution): text Phone Number (optional, enter "-" to skip): text Country: country I intend to use this dataset for: type: select options: - Reward model / preference training - Model evaluation or benchmarking - Annotation-quality research - label: Other value: other extra_gated_button_content: Request access configs: - config_name: audio default: true data_files: - split: train path: data/audio/train-*.parquet - split: test path: data/audio/test-*.parquet - config_name: pairs data_files: - split: train path: data/pairs/train.parquet - split: test path: data/pairs/test.parquet - config_name: responses data_files: - split: train path: data/responses/train.parquet - split: test path: data/responses/test.parquet - config_name: prompts data_files: - split: train path: data/prompts/train.parquet - split: test path: data/prompts/test.parquet - config_name: models data_files: data/models/models.parquet --- Datapoint Audio Bench — 315K votes across 15 models # Text-to-speech human preferences: 315K votes across 15 models This gated dataset contains the evaluation record behind **Datapoint Audio Bench**: **315,000 eligible pairwise votes** comparing 15 text-to-speech models in a complete round-robin over 300 English prompts. The prompt set covers eight practical voice-agent categories, and every generated sample is included as a typed audio record. The source evaluation collected 357,651 completed responses. The published benchmark excluded 41,873 trust-voided responses and deterministically capped 778 overserved responses, leaving the exact 315,000 votes used by the leaderboard published on 2026-09-01. The `responses` config contains only those score-bearing responses; voided and over-cap responses are not included. Built on the [Datapoint](https://trydatapoint.com) annotation platform. ## Key features - **Complete comparison design.** Every one of the 105 model pairings is evaluated on all 300 prompts, for 31,500 comparison cells. - **Audio previews.** The default `audio` config stores all 4,500 source FLAC renders as typed Hugging Face `Audio` features. Each render is stored once rather than duplicated across every comparison in which it appears. The card below also provides direct previews for authorized users when the private-repository Dataset Viewer is unavailable on the owner's Hub plan. - **Operational voice-agent prompts.** The categories cover transactional readouts, empathy and de-escalation, names and spell-outs, instructions and troubleshooting, repairs and disfluency, greetings and brand delivery, policy disclosure, and scheduling and escalation. - **Blind comparison.** Annotators received two unlabeled audio candidates and selected the preferred delivery. The schedule balances which model occupies the first candidate slot across prompts. - **Frozen quality evidence.** Every response contains the audio-trust estimate captured at completion time. Raw votes drive the official benchmark; trust-weighted aggregates are provided for sensitivity analysis. ## Prompt categories The benchmark contains 300 prompts across eight practical customer-support categories. Every prompt is evaluated for all 105 model pairs with ten score-bearing responses per pair. | Category | Dataset value | Prompts | |---|---|---:| | Empathy & De-escalation | `empathy_deescalation` | 45 | | Greetings & Brand | `greetings_brand` | 30 | | Instructions & Troubleshooting | `instructions_troubleshooting` | 40 | | Names & Spell-outs | `names_spellouts` | 45 | | Policy & Disclosure | `policy_disclosure` | 30 | | Repairs & Disfluency | `repairs_disfluency` | 35 | | Scheduling & Escalation | `scheduling_escalation` | 30 | | Transactional Readouts | `transactional_readouts` | 45 | | **Total** | | **300** | ## Audio preview The three clips below render the same held-out empathy prompt. They are representative examples for checking playback and are not a claim about model quality. Access remains subject to this repository's approval gate. > I need to share something serious. Your email address was part of last > week's data exposure — your password was not. Here is exactly what we're > doing about it. **Chatterbox HD** **Eleven v3** **GPT-4o mini TTS** ## Dataset structure | Config | Rows | Description | |---|---:|---| | `audio` (default) | 4,500 | One full 48 kHz, mono, 24-bit FLAC render stored as typed audio | | `pairs` | 31,500 | One prompt/model comparison with vote aggregates and audio join keys | | `responses` | 315,000 | One score-bearing human response with anonymized annotator and vote metadata | | `prompts` | 300 | Prompt script, target delivery, persona, pace, and rubric | | `models` | 15 | Stable model identifier and display name | ### audio Each row stores one model output. This is the default config so opening the Dataset Viewer exposes playable examples whenever the repository owner's Hub plan supports a private viewer. The card-level examples above remain available to authorized users without that feature. | Column | Type | Description | |---|---|---| | `audio_key` | string | Stable `{model_id}/{prompt_id}` join key | | `prompt_id`, `prompt_index`, `category` | string/int | Prompt identity and benchmark category | | `model_id` | string | Stable model identifier | | `audio` | audio | Embedded FLAC bytes rendered as a player by the Hub | | `sample_rate`, `channels`, `bit_depth` | int | Source format metadata | | `total_samples`, `duration_seconds`, `byte_size` | numeric | Audio size and duration metadata | | `sha256` | string | SHA-256 digest of the exact FLAC bytes | ### pairs Each row is one comparison. `audio_a_key` and `audio_b_key` join to the `audio` config without embedding the same FLAC many times. | Column | Type | Description | |---|---|---| | `pair_key` | string | Stable comparison identifier | | `category`, `prompt_id` | string | Prompt and category identity | | `model_a`, `model_b` | string | Models assigned to the two source candidate slots | | `audio_a_key`, `audio_b_key` | string | Join keys into the `audio` config | | `votes_a`, `votes_b` | int | Eligible votes used by the published leaderboard | | `label_a`, `label_b` | float | Preference fractions; 0.5/0.5 when no eligible vote exists | | `trust_weighted_votes_a`, `trust_weighted_votes_b` | float | Eligible vote totals weighted by frozen audio trust | | `winner` | string | `a`, `b`, or `tie` | | `num_votes` | int | Total eligible votes for the comparison, capped at 10 | ### responses Each row is one response used by the published score. A pair has at most ten rows, exactly matching `votes_a + votes_b` in `pairs`; voided and over-cap responses are omitted. Raw account and response IDs are not released. `annotator` is a salted hash that is stable only inside this export. | Column | Type | Description | |---|---|---| | `pair_key` | string | Joins to `pairs` | | `prompt_id`, `category` | string | Prompt identity and benchmark category | | `chosen` | string | `a` or `b` for a valid selection | | `annotator` | string | Export-local salted annotator hash | | `trust_score` | float | Frozen audio-trust estimate used for sensitivity weighting | | `time_taken_ms`, `completed_at`, `country` | mixed | Response timing and country snapshot | ## Splits The `audio`, `pairs`, `responses`, and `prompts` configs share a deterministic, prompt-level train/test split. Thirty prompts are held out with stratification across the eight categories. No prompt or audio render crosses the split, so a reward model can be trained on `train` and evaluated on `test` without prompt leakage. `models` is a reference table with one `train` split. ## Usage Approved gated access and an authenticated Hugging Face token are required. ```python from datasets import load_dataset audio = load_dataset( "datapointai/text-to-speech-human-preferences-315k", "audio", split="train", token=True, ) sample = audio[0] print(sample["model_id"], sample["prompt_id"]) print(sample["audio"]) ``` Load comparison aggregates and join them to audio by key: ```python pairs = load_dataset( "datapointai/text-to-speech-human-preferences-315k", "pairs", split="train", token=True, ) audio_by_key = {row["audio_key"]: row["audio"] for row in audio} pair = pairs[0] audio_a = audio_by_key[pair["audio_a_key"]] audio_b = audio_by_key[pair["audio_b_key"]] print(pair["model_a"], "vs", pair["model_b"], "->", pair["winner"]) ``` For metadata-only iteration without decoding audio: ```python from datasets import Audio audio = audio.cast_column("audio", Audio(decode=False)) ``` ## Benchmark methodology The official category boards fit a deterministic Bradley–Terry model to raw eligible votes and transform the fitted strengths to Elo. Prompt-cluster bootstrap samples provide score uncertainty and rank spread. The Overall board combines the eight published category boards with equal category weight. The official fit does not use `trust_weighted_votes_*`. Those fields support a sensitivity analysis against the frozen audio-trust estimates. ## Leaderboard Elo ratings come from the deterministic Bradley–Terry fit on the 315,000 published responses. The chart shows the Overall board published on 2026-09-01, combining all eight categories with equal weight. Customer support audio model Elo rankings — 15 models ranked by Elo score ## Intended use Use this dataset to: - train or evaluate audio preference and reward models; - study human preference aggregation and quality filtering; - compare text-to-speech systems by prompt category; - reproduce or audit the published benchmark inputs; and - evaluate reranking or best-of-N selection methods for speech generation. ## Limitations and responsible use - The benchmark measures preference on this fixed English prompt set; it does not establish universal speech quality, accessibility, safety, or language coverage. - Model folder names are the stable identities supplied with the source benchmark. Exact provider revisions and generation parameters were not available and must not be inferred from the audio. - Human preference can encode demographic and cultural bias. Category or overall rank should not be treated as a guarantee for every listener or use case. - Do not attempt to re-identify annotators from the export-local hashes, timestamps, or country fields. - Synthetic voices can be misused for impersonation. Follow applicable law, provider terms, disclosure requirements, and consent expectations. ## License Votes, prompts, and Datapoint-authored metadata are provided under **CC-BY-4.0**. Audio files are outputs of the listed third-party models and are included for private evaluation and research; rights and usage restrictions for those outputs remain governed by the applicable model-provider terms.