---
language:
- fr
license: apache-2.0
task_categories:
- text-to-speech
tags:
- french
- tts
- conversational
- multi-speaker
- fintech
- ecommerce
- healthcare
size_categories:
- 100K<n<1M
---

# French Conversational TTS Dataset

## Dataset Description

This dataset contains high-fidelity French text-to-speech audio clips generated using Mistral's **Voxtral Mini TTS** model (`voxtral-mini-tts-2603`). It covers three B2B industry verticals with balanced male/female speaker distribution.

### Verticals

| Vertical | Description |
|---|---|
| `fintech_banking` | Banking operations, account inquiries, fraud alerts, investments, customer service |
| `ecommerce_logistics` | Order tracking, delivery notifications, returns, product info, promotions |
| `healthcare_medical` | Appointments, prescriptions, test results, procedures, health advice, insurance |

### Speaker Distribution

- **6 Female voices** (French native - Marie variants: neutral, happy, excited, curious, sad, angry)
- **6 Male voices** (Multilingual Voxtral - Paul & Oliver variants)
- Strict 50/50 gender rotation across all clips

### Format

- **Audio**: MP3 files organized as `data/audio/{GENDER}_{VERTICAL}/clip_{id}.mp3`
- **Metadata**: JSONL files in `metadata/` with fields: `file_path`, `text`, `speaker_id`, `gender`, `vertical`

### Data Quality

- All text is pre-processed: numbers, acronyms, and symbols converted to spoken French form
- Clip length: 15-40 words per audio file
- No markdown, emojis, or special punctuation in text content