How to Create a High-Quality Dataset

A dataset is a collection of clean, high-quality audio files of a single speaker that Applio uses to train a voice model. The quality of your dataset is the single most important factor in achieving good training results. A high-quality dataset consists of clear, consistent, and noise-free audio.

This guide will walk you through the process of creating a great dataset.

Dataset Creation Workflow

Step 1: Source Your Audio

First, you need to collect audio of the person or character you want to create a voice model for.

Best Practices for Audio Sourcing

• Audio Format: Your audio files should be in a lossless format like WAV or FLAC.
• Duration: Aim for a total duration of 10 to 30 minutes. Less than 10 minutes may not be enough for good results, and more than 30 minutes can lead to diminishing returns.
• Variety: The dataset should include a wide range of tones, pitches, and emotions to create a versatile voice model.
• Quality: The audio should be as clean as possible, with minimal background noise, reverb, or other artifacts.

Step 2: Isolate the Vocals

If your audio contains background music or other sounds, you’ll need to isolate the vocals. For a detailed guide on how to do this, please see our audio isolation guide.

Step 3: Clean and Process Your Audio with Audacity

Once you have your vocal recordings, it’s time to clean and process them using a free audio editor like Audacity.

Noise Reduction

Noise reduction helps to remove unwanted background noise from your recordings.

1. In Audacity, select a small portion of your audio that contains only background noise.
2. Go to Effect > Noise Removal and Repair > Noise Reduction.
3. Click Get Noise Profile.
4. Now, select the entire audio track.
5. Go back to the Noise Reduction effect and click OK to apply it.

Noise Gate

A noise gate is used to silence parts of the audio that are below a certain volume threshold. This is great for removing low-level noise between words and sentences.

1. Select your entire audio track.
2. Go to Effect > Gating > Noise Gate.
3. Apply the recommended settings as shown in the image below. These settings are a good starting point, but you may need to adjust them based on your audio.

Truncate Silence

This effect removes long periods of silence from your audio, which helps to create a more concise dataset.

1. Select your entire audio track.
2. Go to Effect > Truncate / Silence > Truncate Silence.
3. Apply the recommended settings as shown in the image below.

Caution

Before truncating silence, make sure your audio contains only speech. Manually remove any non-speech sounds like coughs, breaths, or clicks.

Step 4: Export Your Dataset

Once you’re happy with your audio, it’s time to export it.

1. Go to File > Export > Export as WAV (or FLAC).
2. Choose a location to save your file.
3. Ensure the format is set to WAV (Microsoft) signed 16-bit PCM or FLAC.
4. Click Export.

WAV Export Settings	FLAC Export Settings

Note

You can export your dataset as a single audio file or as multiple smaller files. If you choose to split your audio, aim for each file to be between 10 and 15 seconds long.

Your dataset is now ready for training!

---

Recommended Audio Editing Plugins

While Audacity is a powerful tool on its own, you can extend its functionality with plugins. These plugins work with Audacity, FL Studio, and other DAWs.

• T-De-Esser: A de-esser is a tool that reduces sibilance, which is the harsh “s” sound in speech. This is a must-have for creating clean vocal recordings.
• ReaPlugs: A suite of powerful plugins from the creators of the Reaper DAW. It includes a more advanced noise gate, EQ, and compressor.
• Auburn Sounds Renegate: A sophisticated noise gate plugin that gives you more control than Audacity’s built-in noise gate.