---
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
language:
- fr
pretty_name: Siwis
size_categories:
- 1K<n<10K
---

# SIWIS French Speech Synthesis Database

This README provides a concise description of the dataset, including its structure, file naming conventions, and known labeling issues. Additionally, suggestions for potential improvements are outlined in the TODO section.  

The dataset is distributed under the [Creative Commons Attribution 4.0 International (CC BY 4.0) license](https://creativecommons.org/licenses/by/4.0/), permitting its use for any purpose.  

For more details about the database design and recording process, please refer to:  
Pierre-Edouard Honnet, Alexandros Lazaridis, Philip N. Garner, and Junichi Yamagishi, *The SIWIS French Speech Synthesis Database – Design and recording of a high-quality French database for speech synthesis,* Idiap-Internal-RR-06-2017.

---

## Dataset Contents  

The dataset is organized into the following components:  

- **`wavs`**: Segmented audio files.  
  - Note: Part 4 contains unsegmented data, and Part 5 is partially segmented.  
- **`text`**: Text prompts corresponding to the audio files.  
- **`lists`**: Recommended training, development, and test sets.  
- **`labs`**: Aligned HTS (Hidden Markov Model-based Text-to-Speech) labels.  
- **`emph_labs`**: Aligned HTS labels with emphasis markers on emphasized words.  
  - `P:0`: Neutral words  
  - `P:1`: Emphasized words  
- **`other`**:  
  - Consolidated prompt files.  
  - Question file for use during HTS training with the provided labels.  

---

## File Naming Conventions  

### Parts 1, 2, 3, and 5  
Files are named using the format:  
`[style]_[type]_[session]_[id].[ext]`  

Where:  
- **`style`**: Speech style  
  - `neut`: Neutral  
  - `emph`: Emphasized  
  - `expr`: Expressive  
- **`type`**: Content type  
  - `parl`: Parliament  
  - `book`: Book  
  - `siwis`: SIWIS prompts  
  - `sus`: SUS (Semantically Unpredictable Sentences)  
  - `chap`: Chapter  
- **`session`**: Two-digit session identifier.  
- **`id`**: Four-digit unique identifier.  
- **`ext`**: File extension (`wav`, `txt`, `lab`).  

### Part 4  
Files use the format:  
`chap_full.[ext]`  

---

## Additional Notes  

- **Speakers**:  
  - Parts 1 to 4 feature recordings from the same main speaker.  
  - Part 5 is recorded by a different speaker.  
- **Labels**:  
  - Parts 1 to 3 include aligned labels.  
  - Parts 4 and 5 lack labels due to their unsegmented nature.  

---

## Known Issues  

1. The `[s]` sound in "c'est" is frequently replaced with `[k]`, affecting label accuracy.  
2. In cases where multiple words are emphasized within a sentence, some emphasized words may not be labeled correctly.  

---

## TODO  

To improve the dataset:  

- Expand labels to include relative position information for emphasized words. For example, indicate the distance of an emphasized word relative to the current word.  

---

For any questions or further clarification, please contact the dataset maintainers.