Command-Line Interface

Applio provides a command-line interface through core.py that allows you to perform all major operations without the Gradio web UI. This is useful for automation, headless servers, cloud environments, and advanced scripting.

The web server (app.py) also accepts a number of startup arguments — see Application Server Arguments below.

Usage

python core.py <command> [options]
python app.py [options]

Global Options

Option	Description
-h, --help	Show help message and exit

---

Inference Commands

infer

Convert a single audio file using a voice model.

python core.py infer --input_path <audio> --output_path <output> --pth_path <model> [options]

Required Arguments:

Option	Description
--input_path	Path to the input audio file
--output_path	Path to save the converted audio
--pth_path	Path to the model .pth file
--index_path	Path to the .index file

Pitch & Voice Arguments:

Option	Type	Default	Description
--pitch	int (-24..24)	0	Pitch adjustment in semitones
--index_rate	float (0..1)	0.75	Search feature ratio — how much the index influences output
--volume_envelope	float (0..1)	1	RMS mix rate — blends output volume with original
--protect	float (0..0.5)	0.5	Protect voiceless consonants and breathing sounds
--f0_method	str	"rmvpe"	Pitch extraction method (crepe, crepe-tiny, rmvpe, fcpe, hybrid)
--sid	int	0	Speaker ID for multi-speaker models

Processing Arguments:

Option	Type	Default	Description
--split_audio	bool	False	Split long audio into segments for processing
--f0_autotune	bool	False	Apply autotune to the output pitch
--f0_autotune_strength	float (0..1)	1.0	Autotune intensity
--proposed_pitch	bool	False	Enable proposed pitch correction
--proposed_pitch_threshold	float (50..1200)	155.0	Threshold for proposed pitch
--clean_audio	bool	False	Apply noise reduction to the output
--clean_strength	float (0..1)	0.7	Noise reduction intensity
--export_format	str	"WAV"	Output format (WAV, MP3, FLAC, OGG, M4A)
--embedder_model	str	"contentvec"	Embedder model (contentvec, spin, spin-v2, chinese-hubert, japanese-hubert, korean-hubert, custom)
--embedder_model_custom	str	None	HuggingFace model ID when embedder_model is custom
--formant_shifting	bool	False	Enable formant shifting
--formant_qfrency	float	1.0	Formant shift quefrency
--formant_timbre	float	1.0	Formant shift timbre
--post_process	bool	False	Enable post-processing effects chain

Post-Processing Effects:

Each effect has a boolean toggle and one or more parameter flags. Set the toggle to enable the effect, then configure it with the parameter flags.

Effect Toggle	Parameter Flags	Defaults
--reverb	--reverb_room_size --reverb_damping --reverb_wet_gain --reverb_dry_gain --reverb_width --reverb_freeze_mode	0.5 each
--pitch_shift	--pitch_shift_semitones	0.0
--limiter	--limiter_threshold --limiter_release_time	-6 / 0.01
--gain	--gain_db	0.0
--distortion	--distortion_gain	25
--chorus	--chorus_rate --chorus_depth --chorus_center_delay --chorus_feedback --chorus_mix	1.0 / 0.25 / 7 / 0.0 / 0.5
--bitcrush	--bitcrush_bit_depth	8
--clipping	--clipping_threshold	-6
--compressor	--compressor_threshold --compressor_ratio --compressor_attack --compressor_release	0 / 1 / 1.0 / 100
--delay	--delay_seconds --delay_feedback --delay_mix	0.5 / 0.0 / 0.5

batch_infer

Convert multiple audio files in a directory.

python core.py batch_infer --input_folder <folder> --output_folder <folder> --pth_path <model> [options]

Option	Alias	Description	Default
--input_folder		Path to folder containing input audio files	Required
--output_folder		Path to folder for output files	Required
--pth_path		Path to the model .pth file	Required
--index_path		Path to the .index file	Required
All inference options		Same as infer command	Same as infer

tts

Generate speech from text using Edge TTS, optionally passing through voice conversion.

python core.py tts --tts_text <text> --tts_voice <voice> --output_tts_path <file> --output_rvc_path <file> --pth_path <model> [options]

Option	Type	Default	Description
--tts_text	str	Required	Text to synthesize
--tts_file	str	None	Path to a .txt file with the text
--tts_voice	str	Required	TTS voice name
--tts_rate	int (-100..100)	0	Speech rate adjustment
--output_tts_path	str	Required	Path to save the raw TTS audio
--output_rvc_path	str	Required	Path to save the voice-converted audio
--pth_path	str	Required	Path to voice model .pth
--index_path	str	Required	Path to .index file
--pitch	int (-24..24)	0	Pitch adjustment
--index_rate	float (0..1)	0.3	Search feature ratio
--volume_envelope	float (0..1)	1	RMS mix rate
--protect	float (0..0.5)	0.33	Protect voiceless consonants
--f0_method	str	"rmvpe"	Pitch extraction method
--split_audio	bool	False	Split audio for processing
--f0_autotune	bool	False	Enable autotune
--f0_autotune_strength	float (0..1)	1.0	Autotune strength
--proposed_pitch	bool	False	Proposed pitch correction
--proposed_pitch_threshold	float (100..500)	155.0	Pitch threshold
--clean_audio	bool	False	Enable noise reduction
--clean_strength	float (0..1)	0.7	Noise reduction strength
--export_format	str	"WAV"	Output format
--embedder_model	str	"contentvec"	Embedder model
--embedder_model_custom	str	None	Custom HuggingFace embedder

---

Training Commands

preprocess

Preprocess a dataset for training. Slices audio, removes silence, applies noise reduction, and resamples.

python core.py preprocess --model_name <name> --dataset_path <path> --sample_rate <rate> [options]

Option	Type	Default	Description
--model_name	str	Required	Name for the model
--dataset_path	str	Required	Path to the dataset folder
--sample_rate	int	Required	Target sample rate (32000, 40000, 48000)
--cpu_cores	int	None	Number of CPU threads to use
--cut_preprocess	str	"Automatic"	Slicing method (Skip, Simple, Automatic)
--process_effects	bool	False	Apply audio effects during preprocessing
--noise_reduction	bool	False	Apply noise reduction
--noise_reduction_strength	float (0..1)	0.7	Noise reduction intensity
--chunk_len	float (0.5..5)	3.0	Slice length in seconds
--overlap_len	float (0..0.4)	0.3	Overlap between slices
--normalization_mode	str	"none"	Peak normalization (none, pre, post)

extract

Extract features (F0 and embeddings) from a preprocessed dataset.

python core.py extract --model_name <name> --sample_rate <rate> [options]

Option	Type	Default	Description
--model_name	str	Required	Name of the model
--sample_rate	int	Required	Sample rate (32000, 40000, 44100, 48000)
--f0_method	str	"rmvpe"	Pitch extraction method
--embedder_model	str	"contentvec"	Embedder model
--embedder_model_custom	str	None	Custom HuggingFace embedder
--cpu_cores	int	None	CPU threads
--gpu	str	"-"	GPU index to use (e.g. "0")
--include_mutes	int (0..10)	2	Number of mute frames to include

train

Train a voice model.

python core.py train --model_name <name> --sample_rate <rate> [options]

Option	Type	Default	Description
--model_name	str	Required	Name of the model
--sample_rate	int	Required	Sample rate (32000, 40000, 48000)
--vocoder	str	"HiFi-GAN"	Vocoder type (HiFi-GAN, MRF HiFi-GAN, RefineGAN)
--total_epoch	int (1..10000)	1000	Total training epochs
--save_every_epoch	int (1..100)	Required	Save checkpoint every N epochs
--batch_size	int (1..50)	8	Training batch size
--gpu	str	"0"	GPU index for training
--pretrained	bool	True	Use default pre-trained model
--custom_pretrained	bool	False	Use a custom pre-trained model
--g_pretrained_path	str	None	Path to custom pre-trained generator
--d_pretrained_path	str	None	Path to custom pre-trained discriminator
--index_algorithm	str	"Auto"	Index algorithm (Auto, Faiss, KMeans)
--checkpointing	bool	False	Enable gradient checkpointing to save VRAM
--save_only_latest	bool	False	Only keep the latest checkpoint
--save_every_weights	bool	True	Save model weights at every checkpoint
--cleanup	bool	False	Clean up temporary files after training
--cache_data_in_gpu	bool	False	Cache training data in GPU memory

index

Generate a FAISS or KMeans index from extracted features.

python core.py index --model_name <name>

Option	Type	Default	Description
--model_name	str	Required	Name of the trained model
--index_algorithm	str	"Auto"	Index algorithm (Auto, Faiss, KMeans)

---

Model Utilities

model_information

Display metadata about a trained model.

python core.py model_information --pth_path <path>

Option	Type	Description
--pth_path	str	Path to the .pth model file (Required)

Output includes: version, sample rate, whether F0 was used, embedder name, epoch count, step count.

model_blender

Blend two model weights into a new model using linear interpolation.

python core.py model_blender --model_name <name> --pth_path_1 <model_a> --pth_path_2 <model_b> [options]

Option	Type	Default	Description
--model_name	str	Required	Name for the blended model
--pth_path_1	str	Required	Path to first model .pth
--pth_path_2	str	Required	Path to second model .pth
--ratio	float (0..1)	0.5	Blend ratio (0 = all model A, 1 = all model B)

tensorboard

Launch TensorBoard for monitoring training progress.

python core.py tensorboard

No arguments.

---

Download & Setup

download

Download a model from a URL (HuggingFace, Google Drive, or direct link).

python core.py download --model_link <url>

Option	Type	Description
--model_link	str	Download URL for the model (Required)

prerequisites

Download and install required prerequisites (pretrained models, predictors, embedders).

python core.py prerequisites [options]

Option	Type	Default	Description
--pretraineds_hifigan	bool	True	Download pre-trained HiFi-GAN models
--models	bool	True	Download predictor models
--exe	bool	True	Download bundled executables (ffmpeg, etc.)

---

Analysis

audio_analyzer

Analyze an audio file and display spectral information.

python core.py audio_analyzer --input_path <audio>

Option	Type	Description
--input_path	str	Path to the audio file to analyze (Required)

---

Application Server Arguments

These arguments are passed to app.py when starting the web interface:

python app.py [options]

Option	Description
--share	Generate a public Gradio share URL (e.g., https://xxxxx.gradio.live)
--port <num>	Specify the server port (default: 6969)
--server-name <host>	Bind to a specific host address (default: 127.0.0.1)
--open	Automatically open the browser on launch
--client	Enable realtime client mode (mounts FastAPI endpoint and injects audio client JS)

Port conflict handling: If the specified port is already in use, Applio automatically decrements the port number and retries until it finds an available port.

---

Examples

Basic inference with a model:

python core.py infer --input_path my_speech.wav --output_path converted.wav --pth_path logs/MyModel/G_1234.pth --index_path logs/MyModel/MyModel.index

Inference with pitch adjustment and CREPE:

python core.py infer --input_path input.wav --output_path output.wav --pth_path logs/MyModel/G_1000.pth --index_path logs/MyModel/MyModel.index --pitch -3 --f0_method crepe --embedder_model contentvec

Batch inference:

python core.py batch_infer --input_folder ./input --output_folder ./output --pth_path logs/MyModel/G_1234.pth --index_path logs/MyModel/MyModel.index

Full training pipeline:

python core.py preprocess --model_name MyVoice --dataset_path assets/datasets/MyVoice --sample_rate 40000
python core.py extract --model_name MyVoice --sample_rate 40000 --f0_method rmvpe --embedder_model contentvec
python core.py train --model_name MyVoice --sample_rate 40000 --f0_method rmvpe --embedder_model contentvec --batch_size 8 --save_every_epoch 50 --total_epoch 300
python core.py index --model_name MyVoice

TTS with voice conversion:

python core.py tts --tts_text "Hello, this is a test." --tts_voice "en-US-JennyNeural" --output_tts_path tts_output.wav --output_rvc_path converted.wav --pth_path logs/MyModel/G_1234.pth --index_path logs/MyModel/MyModel.index

Blend two models:

python core.py model_blender --model_name Blended --pth_path_1 logs/ModelA/G_100.pth --pth_path_2 logs/ModelB/G_200.pth --ratio 0.6

Inspect model metadata:

python core.py model_information --pth_path logs/MyModel/G_1234.pth

Start the web UI with a public share link:

python app.py --share --port 7860