Skip to main content
Matcha-TTS expects audio datasets in a specific format for training. This guide covers the dataset structure requirements and how to prepare your data.

​
Dataset Structure

Matcha-TTS follows the LJSpeech dataset format. Your dataset should be structured as: 
data/
└── YourDataset-1.0
    ├── metadata.csv
    ├── train.txt
    ├── val.txt
    ├── test.txt (optional)
    └── wavs/
        ├── audio001.wav
        ├── audio002.wav
        └── ...

​
Filelist Format

​
Single-Speaker Datasets

For single-speaker datasets (like LJSpeech), create filelist files with the following format: 
path/to/audio.wav|Transcription text here
path/to/audio2.wav|Another transcription
Each line contains:
  • Audio file path (relative or absolute)
  • Pipe separator |
  • Text transcription
Example (train.txt): 
data/LJSpeech-1.1/wavs/LJ001-0001.wav|Printing, in the only sense with which we are at present concerned.
data/LJSpeech-1.1/wavs/LJ001-0002.wav|differs from most if not from all the arts and crafts.

​
Multi-Speaker Datasets

For multi-speaker datasets (like VCTK), add a speaker ID between the path and text: 
path/to/audio.wav|speaker_id|Transcription text here
Each line contains:
  • Audio file path
  • Pipe separator |
  • Speaker ID (integer)
  • Pipe separator |
  • Text transcription
Example: 
data/vctk/wav48/p225/p225_001.wav|0|Please call Stella.
data/vctk/wav48/p226/p226_001.wav|1|Please call Stella.

​
Audio Requirements

Your audio files must meet these specifications:
All audio files must have the same sample rate as specified in your configuration file (default: 22050 Hz).
  • Format: WAV (uncompressed PCM)
  • Sample Rate: 22050 Hz (configurable)
  • Channels: Mono (single channel)
  • Bit Depth: 16-bit recommended

​
Preparing LJSpeech Dataset

1

Download the dataset

Download LJSpeech from the official website.
wget https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2
2

Extract the archive

Extract to your data directory:
tar -xjf LJSpeech-1.1.tar.bz2 -C data/
This creates data/LJSpeech-1.1/ with the required structure.
3

Create train/val splits

Split metadata.csv into training and validation sets. The LJSpeech dataset contains 13,100 audio clips.You can use the NVIDIA Tacotron 2 repository’s file lists as described in item 5 of their setup guide, or create your own:
# Example Python script to create splits
import random

with open('data/LJSpeech-1.1/metadata.csv', 'r') as f:
    lines = f.readlines()

random.shuffle(lines)
split_idx = int(len(lines) * 0.95)  # 95% train, 5% val

# Convert format: LJ001-0001|text|normalized -> path|text
train_lines = [f"data/LJSpeech-1.1/wavs/{line.split('|')[0]}.wav|{line.split('|')[1]}\n" 
               for line in lines[:split_idx]]
val_lines = [f"data/LJSpeech-1.1/wavs/{line.split('|')[0]}.wav|{line.split('|')[1]}\n" 
             for line in lines[split_idx:]]

with open('data/LJSpeech-1.1/train.txt', 'w') as f:
    f.writelines(train_lines)

with open('data/LJSpeech-1.1/val.txt', 'w') as f:
    f.writelines(val_lines)
4

Verify the structure

Ensure your directory looks like:
data/LJSpeech-1.1/
├── metadata.csv
├── README
├── train.txt       # Your training filelist
├── val.txt         # Your validation filelist
└── wavs/
    ├── LJ001-0001.wav
    ├── LJ001-0002.wav
    └── ... (13,100 files)

​
Preparing Custom Datasets

1

Organize audio files

Place all audio files in a wavs/ directory or similar structure.
2

Create transcriptions

Create filelist files (train.txt, val.txt) with audio paths and transcriptions in the format described above.
3

Verify audio format

Ensure all audio files are:
  • WAV format
  • Same sample rate (e.g., 22050 Hz)
  • Mono channel
You can use sox or ffmpeg to convert:
# Convert all files to 22050 Hz mono WAV
for file in wavs/*.wav; do
  ffmpeg -i "$file" -ar 22050 -ac 1 "converted/$file"
done
4

Split data

Create separate train/validation splits. A typical split is 95% training, 5% validation.
Ensure no overlap between training and validation sets to avoid data leakage.

​
Text Preprocessing

Matcha-TTS includes several text cleaners for preprocessing:
  • english_cleaners - Basic English text cleaning
  • english_cleaners2 - Advanced English cleaning (recommended for LJSpeech)
The cleaners handle:
  • Number expansion (“123” → “one hundred twenty three”)
  • Abbreviation expansion (“Dr.” → “Doctor”)
  • Lowercase conversion
  • Punctuation normalization
Text cleaning is configured in your dataset YAML file: 
cleaners: [english_cleaners2]

​
Validation Checklist

Before training, verify:
  • All audio files are in WAV format with correct sample rate
  • Filelist format matches your speaker configuration (single/multi-speaker)
  • No missing audio files referenced in filelists
  • Train/validation splits are created
  • No overlap between train and validation sets
  • All transcriptions are accurate
  • File paths in filelists are correct (relative or absolute)

​
Next Steps

Once your dataset is prepared:
  1. Configure your dataset for training
  2. Follow the Training Guide to start training

Build docs developers (and LLMs) love

Get started for freeTalk to us