Glossary of Terms

A

Autotune

A post-processing effect that adjusts the pitch of the output audio to the nearest semitone. Useful for singing conversions. Applio uses a configurable autotune strength to control the intensity of the effect.

B

Batch Size

The number of training samples processed simultaneously. Larger batch sizes can lead to more stable training but require more GPU VRAM. Common values range from 4 to 24 depending on GPU memory.

Bitcrush

An audio effect that reduces the bit depth or sample rate of audio, intentionally creating a lo-fi, retro sound.

C

Checkpoint

A saved copy of the model’s weights at a specific point during training. Checkpoints are saved at regular intervals (controlled by the “Save Every Epoch” setting) and can be used to resume training or extract a final model.

Chorus

An audio effect that creates a thickened, ensemble-like sound by mixing the original signal with one or more slightly detuned copies.

Clipping

An audio effect that intentionally distorts the signal by cutting off peaks above a certain threshold.

Compressor

An audio effect that reduces the dynamic range of audio by attenuating loud parts and/or amplifying quiet parts.

CREPE

A pitch extraction algorithm that uses a convolutional neural network trained on a large dataset of pitch-labeled audio. Stands for “Convolutional REpresentation for Pitch Estimation.” High accuracy but computationally intensive.

CUDA

NVIDIA’s parallel computing platform that allows software to use NVIDIA GPUs for general-purpose processing. Applio uses CUDA to accelerate training and inference.

D

DDP (Distributed Data Parallel)

A PyTorch technique for training models across multiple GPUs simultaneously, automatically dividing batches across available devices.

Dataset

A collection of audio files used to train a voice model. A good dataset consists of clean, isolated vocals of a single speaker, typically 10-30 minutes in total duration.

Delay

An audio effect that creates echoes by repeating the audio signal after a period of time.

Discriminator (D)

In a GAN (Generative Adversarial Network), the discriminator is a neural network that learns to distinguish between real audio and audio generated by the generator. During training, it competes with the generator, pushing it to produce more realistic outputs.

E

Embedder

A pre-trained neural network (typically based on HuBERT) that converts raw audio into high-dimensional feature vectors called embeddings. The embedder captures linguistic content, acoustic features, and speaker characteristics. See the Embedders guide for more information.

Embedding

A numerical representation (vector) of audio data produced by an embedder. Embeddings encode the essential features of the audio in a form that the voice conversion model can process.

Epoch

One complete pass through the entire training dataset during model training. Training typically requires 200-400 epochs for good results.

Export Format

The audio file format for output. Supported formats include WAV (lossless), MP3 (lossy compressed), FLAC (compressed lossless), OGG (lossy compressed), and M4A (AAC compressed).

F

F0 (Fundamental Frequency)

The lowest frequency of a periodic waveform, perceived as the pitch of a voice. F0 extraction is the process of measuring the pitch contour of an audio signal.

FAISS

Facebook AI Similarity Search — a library for efficient similarity search and clustering of dense vectors. Used in RVC to build and query the feature index during the retrieval step.

FCPE

A fast and accurate pitch extraction model. Stands for “Fully Convolutional Pitch Estimator.” Provides a good balance between speed and accuracy.

Filter Radius

A parameter that controls the smoothing of the F0 contour. Higher values reduce breathiness and artifacts but may smooth out natural pitch variation.

Formant

A resonant frequency of the human vocal tract that shapes the character of speech sounds. Formant shifting adjusts these frequencies, which is particularly useful for making male-to-female and female-to-male conversions sound more natural.

FP16 / FP32

Floating-point precision formats. FP16 (half-precision) uses 16 bits per number, reducing memory usage and improving speed on compatible hardware. FP32 (single-precision) uses 32 bits per number for maximum accuracy. FP16 is supported on NVIDIA RTX 20-series and newer GPUs.

G

Gain

An audio effect that adjusts the overall volume of the signal.

GAN (Generative Adversarial Network)

A machine learning framework where two neural networks — a generator and a discriminator — compete against each other. The generator creates synthetic data, and the discriminator tries to distinguish real from synthetic. This competition drives both networks to improve.

Generator (G)

In a GAN, the generator is the neural network that creates synthetic audio from input features. The trained generator is what gets extracted as the final model for inference.

H

HiFi-GAN

A high-fidelity neural vocoder (generator) architecture that produces high-quality audio waveforms from mel spectrograms. The standard vocoder in Applio.

Hop Length

The number of audio samples between successive analysis frames during pitch extraction. Smaller values provide more precise pitch tracking but require more computation.

HuBERT

A self-supervised learning model for speech representation, developed by Facebook AI. HuBERT models are commonly used as embedders in voice conversion systems.

I

Index File

A file (.index) containing a FAISS or KMeans index of embeddings extracted from the training dataset. During inference, this index is searched to find the most similar features to blend into the output.

Inference

The process of using a trained model to convert an audio file from one voice to another. Also called voice conversion.

K

KMeans Index

An alternative to FAISS for the feature index. Uses k-means clustering to group similar features. Can be faster to search but may be less accurate than FAISS.

L

Limiter

An audio effect that prevents the signal from exceeding a specified amplitude threshold, preventing clipping and distortion.

M

Mel Spectrogram

A visual representation of audio frequencies over time, scaled to the mel scale (which approximates human pitch perception). The mel spectrogram is the target representation that the generator learns to produce during training.

Model Blending

The process of combining two voice models through linear interpolation of their weights. The blend ratio controls how much each model contributes to the final result.

MP3

A lossy audio compression format that reduces file size by removing perceptually less important audio information. Not recommended for datasets but suitable for sharing converted audio.

N

NSF HiFi-GAN

A variant of HiFi-GAN that incorporates Neural Source-Filter (NSF) modeling for improved pitch control. Better suited for singing voices and extreme pitch shifts.

O

OGG

A lossy audio compression format using Vorbis encoding. Provides better quality than MP3 at similar bitrates.

Overtraining

When a model learns the training data too well, including its noise and imperfections, resulting in poor performance on new data. Monitoring the g/total loss graph in TensorBoard helps identify the optimal stopping point.

P

Pitch

The perceived frequency of a sound. In voice conversion, pitch can be adjusted independently of the voice timbre to change the pitch of the output voice.

Pitch Shift

An audio effect that adjusts the overall pitch by a specified number of semitones.

Pre-trained Model

A base model that has already been trained on a large, diverse dataset. Used as a starting point for training a new voice model, which can significantly reduce training time and improve results, especially with small datasets.

PTH File

A PyTorch checkpoint file containing the model’s weights and training metadata. The .pth file is the primary model file used for both training and inference.

Q

Quefrency

In formant shifting, the quefrency controls the scaling of formant frequencies. Lower values produce a “smaller” vocal tract (closer to female/child), higher values produce a “larger” vocal tract (closer to male).

R

RefineGAN

A generator architecture that iteratively refines the output audio quality through multiple refinement stages. Can produce higher quality results but requires more computation.

Reverb

An audio effect that simulates the natural reverberation of sound in a physical space by adding a series of decaying echoes.

RMVPE

A pitch extraction model based on a ResNet architecture. Stands for “Robust Multi-View Pitch Estimation.” The recommended default in Applio due to its excellent balance of speed and accuracy.

RVC (Retrieval-Based Voice Conversion)

The core technology behind Applio. RVC uses a retrieval mechanism — searching a pre-built index of speaker features during inference — to achieve high-quality voice conversion with good timbre preservation.

S

Sample Rate

The number of audio samples per second, measured in Hz. Common rates are 32kHz (32000 Hz), 40kHz (40000 Hz), and 48kHz (48000 Hz). Higher sample rates capture more high-frequency content but require more processing power.

Search Feature Ratio

A parameter controlling how heavily the model relies on the index during retrieval. Higher values (closer to 1.0) produce more accurate timbre matching but may introduce artifacts. Lower values (closer to 0.0) produce cleaner output with less timbre similarity.

Spectrogram

A visual representation of the spectrum of frequencies in audio as they vary with time. Used in the Audio Analyzer tool to identify noise, artifacts, and frequency distribution in audio files.

Split Audio

An inference option that divides long audio files into smaller segments for processing, then reassembles them. Helps avoid memory issues with very long files.

T

Timbre

The quality or “color” of a sound that distinguishes different voices or instruments, independent of pitch and loudness.

TTS (Text-to-Speech)

The synthesis of spoken audio from written text. Applio’s built-in TTS uses Microsoft Edge TTS and can optionally pipe the output through a voice conversion model.

U

Upscale Audio

An inference option that enhances the quality of low-resolution input audio before conversion.

V

VAD (Voice Activity Detection)

A technology that detects whether a person is speaking in an audio stream. In realtime inference, VAD pauses processing during silence to save computational resources.

VC (Voice Conversion)

The process of transforming the voice characteristics of an audio signal from one speaker to another while preserving linguistic content.

Vocoder

A neural network that converts acoustic features (like mel spectrograms) back into raw audio waveforms. Applio supports HiFi-GAN, NSF HiFi-GAN, and RefineGAN vocoders.

VRAM

Video RAM — the memory available on a GPU. The amount of VRAM directly limits the maximum batch size and model complexity that can be used for training and inference.

W

WAV

An uncompressed audio format (Waveform Audio File Format). The standard format for training datasets due to its lossless nature.