Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/IAHispano/Applio/llms.txt

Use this file to discover all available pages before exploring further.

Beyond inference and training, the Applio CLI includes a set of utility subcommands for inspecting models, merging model weights, downloading models from the internet, setting up the runtime environment, analyzing audio files, and launching TensorBoard. These commands are independent of the training pipeline and can be run at any time from the Applio root directory with the virtual environment activated.

​
model_information

The model_information subcommand prints metadata stored inside a .pth model file — including the model architecture, the epoch at which it was saved, the sample rate it was trained at, and the embedder model used. This is useful for auditing unknown or shared models before using them for inference or blending.

​
Flags

​
--pth_path
string
required
Path to the .pth model file to inspect.

​
Example

python core.py model_information \
  --pth_path logs/MyModel/MyModel.pth
The command prints a summary of the model’s stored configuration to stdout, including fields such as architecture version, training epoch, sample rate, and the embedder used during training.

​
model_blender

The model_blender subcommand fuses two RVC models by linearly interpolating their weights at a configurable ratio. The result is a new model file that combines vocal characteristics from both inputs. This technique — often called model merging or model interpolation — can produce hybrid voices that blend the timbre or style of two different source models.

​
Flags

​
--model_name
string
required
Name for the output blended model. The resulting .pth file will be saved to logs/<model_name>/.
​
--pth_path_1
string
required
Path to the first source .pth model file.
​
--pth_path_2
string
required
Path to the second source .pth model file.
​
--ratio
float
default:"0.5"
Blend ratio between the two models. Range: 0.0–1.0 in steps of 0.1. A value of 0.0 produces a copy of model 1; 1.0 produces a copy of model 2; 0.5 produces an equal blend. Intermediate values interpolate smoothly between the two.

​
Example

python core.py model_blender \
  --model_name BlendedVoice \
  --pth_path_1 logs/ModelA/ModelA.pth \
  --pth_path_2 logs/ModelB/ModelB.pth \
  --ratio 0.5
Both models must have been trained with the same architecture (same sample rate and vocoder settings) for the blend to produce meaningful results. Blending incompatible architectures may fail or produce silent / degraded output.

​
download

The download subcommand downloads a model archive from a direct URL or a Hugging Face model repository link, extracts the contents, and places the resulting files in the logs/ directory so they are immediately available for inference.

​
Flags

​
--model_link
string
required
A direct download URL (e.g., a .zip file) or a Hugging Face repository URL pointing to a model archive. The link must be publicly accessible.

​
Example

python core.py download \
  --model_link https://huggingface.co/YourOrg/YourModel/resolve/main/MyVoice.zip
After the command completes, the model files will be available under logs/ ready for use with infer or tts.
Only download models from trusted sources. Applio executes the download directly — no integrity verification beyond format detection is performed. Review model licenses before using models in production.

​
prerequisites

The prerequisites subcommand downloads and installs the files that Applio requires at runtime: official pretrained base model weights, supporting models (e.g., RMVPE), and platform-specific executables. Run this command once after a fresh installation, or again if any required files are missing.

​
Flags

​
--pretraineds_hifigan
boolean
default:"True"
Download the official Applio pretrained HiFi-GAN generator and discriminator weights used as a training starting point. Strongly recommended. Accepts True or False.
​
--models
boolean
default:"True"
Download additional supporting models, including RMVPE and other pitch-extraction or embedding models. Accepts True or False.
​
--exe
boolean
default:"True"
Download required platform-specific executables (e.g., FFmpeg binaries used for audio processing). Accepts True or False.

​
Example

# Download everything (recommended for a fresh install)
python core.py prerequisites \
  --pretraineds_hifigan True \
  --models True \
  --exe True

# Download only the pretrained weights, skip executables
python core.py prerequisites \
  --pretraineds_hifigan True \
  --models True \
  --exe False
If you have already installed Applio’s executables through another method (e.g., your system has FFmpeg in PATH), you can skip the --exe True download to avoid redundancy.

​
audio_analyzer

The audio_analyzer subcommand reads an audio file and prints a summary of its technical properties: sample rate, duration, bit depth, channel count, and RMS signal level. It also generates a waveform and spectrogram visualization saved as a PNG file at logs/audio_analysis.png.

​
Flags

​
--input_path
string
required
Path to the audio file to analyze. Supports common audio formats (WAV, MP3, FLAC, etc.).

​
Outputs

The command prints the following information to stdout:
  • Sample rate (Hz)
  • Duration (seconds)
  • Bit depth
  • Number of channels (mono / stereo)
  • RMS level (dBFS)
It also saves a plot image to logs/audio_analysis.png.

​
Example

python core.py audio_analyzer \
  --input_path audio/sample.wav
Example output: 
Audio info of audio/sample.wav: {'sample_rate': 44100, 'duration': 12.3, ...}
Audio file audio/sample.wav analyzed successfully. Plot saved at: logs/audio_analysis.png
Use audio_analyzer to verify that your training dataset has the expected sample rate and that audio levels are reasonable before running preprocess. Consistently low or clipped RMS levels are common causes of poor model quality.

​
tensorboard

The tensorboard subcommand launches a TensorBoard server pointed at Applio’s logs/ directory. This lets you monitor training loss curves, inspect audio samples written during training, and compare multiple training runs in a browser-based UI.

​
Flags

This subcommand takes no arguments.

​
Example

python core.py tensorboard
After running this command, TensorBoard will start and print a local URL (typically http://localhost:6006) that you can open in a browser. The server will remain running until you interrupt it with Ctrl+C.
TensorBoard data is only available while the logs/ directory contains event files written by one or more training runs. Start a training job first, then open TensorBoard to view live or historical metrics.

Build docs developers (and LLMs) love

Get started for freeTalk to us