Preparing Audio Data

The Data tab turns long recordings of one speaker into a training dataset for fine-tuning. Silero VAD detects the speech, Whisper transcribes it, and the result is cut into clips of a few seconds each.

Audio requirements

Works well:

• A single speaker
• Clean recordings with minimal background noise and no music
• 30+ minutes in total; 1-3 hours is better
• Variety of content and speaking styles

Doesn't work:

• Overlapping speakers
• Music beds, heavy noise or reverb
• Very quiet or distorted audio

Quality matters more than quantity: 30 minutes of clean audio beats 3 hours of noisy audio.

Building a dataset

1. Put your recordings in one folder. Subfolders are searched too, and files can be any length.

   /path/to/recordings/
   ├── episode1.mp3
   ├── episode2.wav
   └── more/chapter3.flac

   Supported formats: WAV, MP3, FLAC, M4A, OGG and Opus.
2. Open the Data tab.
3. Recordings folder: the full path to that folder, e.g. /home/me/recordings or C:\Users\me\recordings.
4. Dataset name: e.g. my_voice (letters, numbers and underscores).
5. Whisper model: openai/whisper-large-v3-turbo (the default) is fast and accurate. Use whisper-large-v3 for the best accuracy, or a smaller model on slow machines.
6. Language (optional): e.g. english. If you leave it empty, the language is detected automatically.
7. Segmentation (optional): Min clip length (default 7 s) and Max clip length (default 20 s).
8. Click Build dataset.

When it's done, the dataset appears under Datasets with its number of clips.

Output format

Datasets are saved in LJSpeech format in data/<name>/:

data/my_voice/
├── wavs/
│   ├── my_voice_000000.wav
│   ├── my_voice_000001.wav
│   └── ...
├── metadata.csv
└── info.json

metadata.csv has one line per clip, in the form filename|text|text:

my_voice_000000.wav|This is the transcribed text.|This is the transcribed text.

You can edit metadata.csv by hand to fix transcription mistakes before training.

Checking quality

Listen to a few random clips in wavs/ and compare them with metadata.csv. Check that:

• each clip starts and ends cleanly, without cut-off words;
• the transcripts match the audio;
• there's no music or second speaker.

If many transcripts are wrong, rebuild with a larger Whisper model or set the language explicitly.

Troubleshooting

• "No speech found in the input files": the audio is too quiet or noisy, or isn't speech. Normalise the volume or use cleaner files.
• Slow: Whisper is slow on CPU. Use a GPU or a smaller Whisper model. You only need to build a dataset once.
• Out of memory: close other GPU apps, or use a smaller Whisper model.

Next step

👉 Finetuning: train a model on your dataset

Copyright and consent

Only train on audio you have the rights to, and only on voices of people who agreed to it.