README.md

---
pretty_name: TTS Voice Direction Benchmark
language:
  - en
task_categories:
  - text-to-speech
tags:
  - text-to-speech
  - voice-direction
  - voice-cloning
  - benchmark
---

# TTS Voice Direction Benchmark

🏆 [Leaderboard](https://breezeblue.ai/benchmarks/tts-voice-direction) | 🛠️ [Evaluation Suite](https://github.com/breezeblue-ai/tts-voice-direction-benchmark)

TTS Voice Direction is a benchmark of **700 reference-conditioned speech
generation tasks**. It evaluates whether a text-to-speech model can preserve a
reference speaker while following a natural-language direction that controls
how a new transcript is performed.

The benchmark emphasizes practical voice acting beyond basic emotion control.
It covers accent, acoustic delivery, vocal events, emotion, physiological
state, communicative intent, role performance, multi-attribute composition,
and temporal variation.

## Dataset At A Glance

- **700** English voice-direction tasks
- **25** synthetic reference voices
- **9** fine-grained direction categories
- **3** macro categories
- One reference audio, reference transcript, direction, and target transcript
  per task
- Four-digit output IDs from `0001` through `0700`

## Taxonomy

| Macro category | Fine-grained category | Cases |
|---|---|---:|
| Foundational Speech Control | Accent | 50 |
| Foundational Speech Control | Acoustic Attributes | 100 |
| Foundational Speech Control | Vocal Events | 50 |
| Situational Voice Acting | Emotion | 150 |
| Situational Voice Acting | Physiological State | 50 |
| Situational Voice Acting | Communicative Intent | 50 |
| Situational Voice Acting | Role | 100 |
| Complex | Composition | 100 |
| Complex | Variation | 50 |

Foundational cases target directly controllable speech properties. Situational
cases describe a high-level state, purpose, or role and require the model to
realize an appropriate performance. Complex cases combine multiple controls or
request a perceptible change over time.

## Reference Voices

All 25 reference voices are synthetic and sourced from the BreezeBlue Voice
Library. Reference files are stored in `reference_audio/` and named
`voice-01.wav` through `voice-25.wav`.

The benchmark rotates directions across the reference set so that each model is
tested on both instruction following and speaker preservation. Reference voice
identity is evaluated separately from direction following.

## Data Format

Each line in `voice_direction_dataset.jsonl` is one task:

```json
{
  "description_id": "0001",
  "language": "en",
  "macro_category": "foundational_speech_control",
  "category": "accent",
  "voice_id": "voice-15",
  "ref_audio_path": "reference_audio/voice-15.wav",
  "ref_audio_text": "Reference transcript...",
  "description": "Use a moderate, consistent General American English accent throughout the line.",
  "transcript": "Target transcript..."
}
```

| Field | Description |
|---|---|
| `description_id` | Stable four-digit task ID |
| `language` | Target transcript language; currently `en` |
| `macro_category` | One of the three aggregation groups |
| `category` | One of the nine fine-grained direction categories |
| `voice_id` | Stable reference-voice identifier |
| `ref_audio_path` | Reference audio path relative to the dataset file |
| `ref_audio_text` | Transcript of the reference audio |
| `description` | Natural-language voice-direction instruction |
| `transcript` | Text to synthesize |

## Running The Benchmark

For every record:

1. Condition the model on `ref_audio_path` and, when required by the model,
   `ref_audio_text`.
2. Use `description` as the voice-direction instruction.
3. Synthesize the exact `transcript`.
4. Save one audio file named after `description_id`, for example `0001.wav`.

Do not use `macro_category` or `category` as additional model inputs. They are
provided only for evaluation and analysis.

## Evaluation

The benchmark reports two complementary metrics:

- **Voice Direction Score (VDS)** is a holistic 1-5 judge score for audible
  fulfillment of the requested direction, requested degree or timing, semantic
  preservation, and coherent execution.
- **Speaker Similarity (SIM)** is cosine similarity between WavLM-Large ECAPA
  embeddings of the generated and reference audio.

Both metrics are aggregated in the same hierarchy: reference voices are
macro-averaged within each fine-grained category, fine-grained categories are
equally averaged within each macro category, and the three macro-category
scores are equally averaged into the overall score.

The self-contained judge prompts, evaluation scripts, speaker-embedding code,
and metric documentation are available in the
[evaluation suite](https://github.com/breezeblue-ai/tts-voice-direction-benchmark).