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.

Every RVC voice conversion begins by extracting a speaker-independent content representation from the input audio. This is the job of the embedder model — a self-supervised speech model (typically a HuBERT variant) that encodes the linguistic content of the source audio while suppressing speaker identity. The resulting embedding is then fed into the RVC synthesizer alongside the target speaker’s learned features from the .pth model file. Choosing the right embedder is critical: the embedder used at inference must match the one used when the model was trained, or the voice conversion quality will degrade noticeably.

​
Available Embedder Models

Applio ships with six built-in embedder options and the ability to load a custom model from disk.
The original HuBERT-based content encoder used by the vast majority of publicly released RVC models. contentvec was trained on English speech and generalises well across languages, making it the safe default for any model whose training embedder is unknown.Best for: General-purpose voice conversion; any model downloaded from the community.CLI value: contentvec
SPIN (Self-supervised Pitch-Invariant Network) is an alternative speech representation model designed to be more robust to pitch variation than standard HuBERT models. It can improve naturalness for certain voices, particularly in singing scenarios.Best for: Models explicitly trained with the spin embedder.CLI value: spin
The second iteration of SPIN, with refined training and improved content disentanglement compared to the original. Use spin-v2 when the target model was trained with spin-v2.Best for: Models explicitly trained with the spin-v2 embedder.CLI value: spin-v2
A HuBERT model fine-tuned on Mandarin Chinese speech data. It captures phonetic distinctions in Chinese more accurately than contentvec, which was primarily trained on English.Best for: Models trained on Mandarin Chinese speech data, or when converting Mandarin Chinese source audio into a Mandarin Chinese target voice.CLI value: chinese-hubert-base
A HuBERT model fine-tuned on Japanese speech data. Japanese phonology (morae, pitch accent, geminate consonants) benefits from a model specifically trained on Japanese audio.Best for: Models trained on Japanese speech.CLI value: japanese-hubert-base
A HuBERT model fine-tuned on Korean speech data.Best for: Models trained on Korean speech.CLI value: korean-hubert-base
Load any compatible HuBERT-architecture model from a local directory on disk. The custom model path is passed separately via --embedder_model_custom.Best for: Experimental or proprietary embedder models not included in the Applio distribution.CLI value: customSee Custom Embedder Usage below for details.

​
How to Choose the Right Embedder

1

Check the model's metadata

Run the model_information command to inspect the training metadata embedded in the .pth file. This will tell you which embedder was used during training:
python core.py model_information --pth_path logs/MyModel/MyModel.pth
Look for the embedder_model field in the output.
2

Match the training embedder

Set --embedder_model to exactly the same value that was used during training. If the metadata does not specify an embedder, assume contentvec — it is the historical default for RVC models.
3

Test and compare

If you are unsure or experimenting with a new model, run a short inference clip with contentvec and your candidate embedder, then compare the outputs. Mismatched embedders typically produce buzzing artifacts, unnatural consonants, or heavily degraded speech intelligibility.
Using a mismatched embedder is one of the most common causes of poor voice conversion quality. A model trained with japanese-hubert-base will produce degraded results when run with contentvec, even if both models share the same architecture.

​
Custom Embedder Usage

To use a model that is not included in Applio’s built-in set:
1

Place your model files

Copy your custom embedder’s model files into the custom embedder directory:
rvc/models/embedders/embedders_custom/
The directory should contain the model weights and any required configuration files expected by the HuBERT loading code.
2

Set the embedder flags

Pass custom as the embedder model and provide the path to your model:
python core.py infer \
  --embedder_model custom \
  --embedder_model_custom rvc/models/embedders/embedders_custom/my_model \
  --input_path audio/input.wav \
  --output_path audio/output.wav \
  --pth_path logs/MyModel/MyModel.pth \
  --index_path logs/MyModel/MyModel.index

​
CLI Example — Language-Specific Embedder

python core.py infer \
  --embedder_model japanese-hubert-base \
  --input_path audio/input.wav \
  --output_path audio/output.wav \
  --pth_path logs/JapaneseModel/model.pth \
  --index_path logs/JapaneseModel/model.index

​
Python API Example

from core import run_infer_script

# Using a language-specific embedder
message, output_path = run_infer_script(
    pitch=0,
    index_rate=0.3,
    volume_envelope=1.0,
    protect=0.33,
    f0_method="rmvpe",
    input_path="audio/input.wav",
    output_path="audio/output.wav",
    pth_path="logs/JapaneseModel/model.pth",
    index_path="logs/JapaneseModel/model.index",
    split_audio=False,
    f0_autotune=False,
    f0_autotune_strength=1.0,
    proposed_pitch=False,
    proposed_pitch_threshold=155.0,
    clean_audio=False,
    clean_strength=0.7,
    export_format="WAV",
    embedder_model="japanese-hubert-base",
    post_process=False,
)

# Using a custom embedder
message, output_path = run_infer_script(
    # ... other params ...
    pitch=0,
    index_rate=0.3,
    volume_envelope=1.0,
    protect=0.33,
    f0_method="rmvpe",
    input_path="audio/input.wav",
    output_path="audio/custom_output.wav",
    pth_path="logs/MyModel/MyModel.pth",
    index_path="logs/MyModel/MyModel.index",
    split_audio=False,
    f0_autotune=False,
    f0_autotune_strength=1.0,
    proposed_pitch=False,
    proposed_pitch_threshold=155.0,
    clean_audio=False,
    clean_strength=0.7,
    export_format="WAV",
    embedder_model="custom",
    embedder_model_custom="rvc/models/embedders/embedders_custom/my_model",
    post_process=False,
)
If you are training your own RVC model, record which embedder you use during the feature extraction step (run_extract_script). Share that information alongside your .pth and .index files so that others can use the correct embedder at inference time.

Build docs developers (and LLMs) love

Get started for freeTalk to us