🎵VIETNAMESE RVC🎵

Simple high-quality and high-performance voice conversion and training.

Full version: Vietnamese-RVC

Music Separation

A simple music separation system can separate into 4 parts: Instruments, Vocals, Main vocals, Backup vocals

Denoise separation

Separate backup vocals

Remove vocal reverb

Enable TTA

High-end processing

Post-processing

Music separation model

Backup separation model

Reverb separation model

Denoise model

Shift

Higher is better quality but uses more resources

1 20

Batch size

The number of samples processed at once. Dividing into batches optimizes the calculation process

1 64

Segments Size

Higher is better quality but uses more resources

32 3072

Aggressiveness

Main body extraction intensity.

1 50

Drop audio here

Url audio

Overlap

Overlap amount between prediction windows

0.25 0.5 0.75 0.99

Window size

A smaller window size will yield higher quality output, but will take longer to process.

320 1024

Hop Length

Analyzing the time transfer window when performing transformations is allowed. The detailed value is compact but requires more calculation

64 8192

Post-processing threshold

The level of post-processing after the music is extracted.

0.1 0.3

Sample rate

NOTE: SOME FORMATS DO NOT SUPPORT RATES ABOVE 48000

8000 11025 12000 16000 22050 24000 32000 44100 48000 88200 96000 176400 192000 352800 384000

Export format

The export format to export the audio file in

wav mp3 flac ogg opus m4a mp4 aac alac wma aiff webm ac3

Input audio path

Output audio folder path

Enter the folder path where the audio will be exported

Audio input

Separated output

Instruments

Original vocal

Main vocal

Backup vocal

Convert Audio

Convert audio using a trained voice model

Clean audio

Autotune

Use separated audio

Using memory-efficient training

Convert original voice

Convert backup voice

Do not merge backup voice

Merge instruments

Pitch

Recommendation: set to 12 to change male voice to female and vice versa

-24 24

Audio cleaning strength

Strength of the audio cleaner for filtering vocals during export

0 1

Select separate files

Drop audio here

Audio input

Model file

Index file

Index strength

Higher values increase strength. However, lower values may reduce artificial effects in the audio

0 1

Probe clusters

The number of clusters the algorithm will examine during approximate nearest neighbor search. Lower values are faster but less accurate, while higher values are more accurate but slower.

1 64

Speaker ID

Speaker ID for the multi-voice model

Export format

The export format to export the audio file in

wav mp3 flac ogg opus m4a mp4 aac alac wma aiff webm ac3

Input audio path

Enter the path to the audio file

Audio output path

Enter the output path (leave it as .wav format; it will auto-correct during conversion)

Extracting pitch using the ONNX model can help improve speed

F0 ONNX

Unlock all pitch extraction methods

Unlock all

F0 method

Method used for data extraction

mangio-crepe-full crepe-full fcpe rmvpe harvest-stonemask pyin hybrid

HYBRID F0 method

Combination of 2 or more different types of extracts

Hop Length

Analyzing the time transfer window when performing transformations is allowed. The detailed value is compact but requires more calculation

64 512

Blending Level of Methods

The blending ratio of pitch values extracted by different methods.

0.1 1

Upload F0 file

F0 File

Blend multiple embedding layers together to achieve better audio quality.

Layer Mixing

Embedders Mode

Extracting embeddings using different models

fairseq onnx transformers

Embedding model

The extracts audio features (Note: Use only Contentvec or HuBERT Base; other models require separate training from scratch).

hubert_base contentvec_base vietnamese_hubert_base japanese_hubert_base korean_hubert_base chinese_hubert_base portuguese_hubert_base spin-v1 spin-v2 custom

Model name

If you have your own model, just upload it and input the name here

Layers to Mix

Lower layers (1-6): Focus on basic acoustic and phonetic details.
Middle layers (7-9): Contain richer phonetic information and work well with linguistic units like phonemes.
Upper layers (10-12): Capture more abstract, high-level information — closer to meaning and speaker-independent content, but with less raw acoustic detail.

1 12

Mix Ratio

Controls how much each layer contributes to the final embedding. Higher values give more weight to higher layers.

0.1 1

Preset file

Save cleanup

Save autotune

Save pitch

Save index impact

Save resampling

Save median filter

Save sound envelope

Save sound protection

Save sound split

Pitch and Formant Shift

File name to save

Upload preset file

Split audio

Pitch and Formant Shift

Automatically propose pitch

Input/output processing

Uses machine learning models to generate additional high-frequency spectrum layers, allowing the output audio to achieve a true sample rate instead of a virtual one created by resampling.

Audio Upscaler

Resample

Resample post-processing to the final sample rate; 0 means no resampling, NOTE: SOME FORMATS DO NOT SUPPORT SPEEDS OVER 48000

0 8000 11025 12000 16000 22050 24000 32000 44100 48000 88200 96000 176400 192000 352800 384000

Proposal pitch threshold

Proposal pitch frequency threshold, for male models used at 155.0 and female models at 255.0

50 1200

Autotune rate

Level of autotuning adjustment

0 1

Filter radius

If greater than three, median filtering is applied. The value represents the filter radius and can reduce breathiness or noise.

0 7

RMS Mix Rate

Determines the blend ratio between the RMS energy of the original voice and the converted voice

0 1

Protect

Protect distinct consonants and breathing sounds to prevent audio tearing and other artifacts. Increasing this value provides comprehensive protection. Reducing it may reduce protection but also minimize indexing effects

0 1

Noise scale

A special adjustment that directly affects audio quality; you should not modify it.

0.1 1

Frequency for Formant Shift

0 16

Timbre for Formant Transformation

0 16

Converted audio

Convert main voice

Convert backup voice

Main voice + Backup voice

Convert original voice

Voice + Instruments

Convert Audio With Voice Activity Detector

Use Voice Activity Detection (VAD) combined with the SpeechBrain model to automatically identify speakers in an audio file. Then, segment and split the audio into smaller individual clips, after which voice conversion is applied using a voice model.

Clean audio

Autotune

Using memory-efficient training

Pitch and Formant Shift

Automatically propose pitch

Input/output processing

Number of voices

Number of voices in the audio

2 8

Model file

Index file

Pitch

Recommendation: set to 12 to change male voice to female and vice versa

-24 24

Index strength

Higher values increase strength. However, lower values may reduce artificial effects in the audio

0 1

Probe clusters

The number of clusters the algorithm will examine during approximate nearest neighbor search. Lower values are faster but less accurate, while higher values are more accurate but slower.

1 64

Speaker ID

Speaker ID for the multi-voice model

Noise scale

A special adjustment that directly affects audio quality; you should not modify it.

0.1 1

Export format

The export format to export the audio file in

wav mp3 flac ogg opus m4a mp4 aac alac wma aiff webm ac3

Input audio path

Enter the path to the audio file

Audio output path

Enter the output path (leave it as .wav format; it will auto-correct during conversion)

Drop audio here

Model file

Index file

Pitch

Recommendation: set to 12 to change male voice to female and vice versa

-24 24

Index strength

Higher values increase strength. However, lower values may reduce artificial effects in the audio

0 1

Probe clusters

The number of clusters the algorithm will examine during approximate nearest neighbor search. Lower values are faster but less accurate, while higher values are more accurate but slower.

1 64

Speaker ID

Speaker ID for the multi-voice model

Noise scale

A special adjustment that directly affects audio quality; you should not modify it.

0.1 1

Extracting pitch using the ONNX model can help improve speed

F0 ONNX

Unlock all pitch extraction methods

Unlock all

F0 method

Method used for data extraction

mangio-crepe-full crepe-full fcpe rmvpe harvest-stonemask pyin hybrid

HYBRID F0 method

Combination of 2 or more different types of extracts

Hop Length

Analyzing the time transfer window when performing transformations is allowed. The detailed value is compact but requires more calculation

64 512

Blending Level of Methods

The blending ratio of pitch values extracted by different methods.

0.1 1

Blend multiple embedding layers together to achieve better audio quality.

Layer Mixing

Embedders Mode

Extracting embeddings using different models

fairseq onnx transformers

Embedding model

The extracts audio features (Note: Use only Contentvec or HuBERT Base; other models require separate training from scratch).

hubert_base contentvec_base vietnamese_hubert_base japanese_hubert_base korean_hubert_base chinese_hubert_base portuguese_hubert_base spin-v1 spin-v2 custom

Model name

If you have your own model, just upload it and input the name here

Layers to Mix

1 12

Mix Ratio

Controls how much each layer contributes to the final embedding. Higher values give more weight to higher layers.

0.1 1

Resample

Resample post-processing to the final sample rate; 0 means no resampling, NOTE: SOME FORMATS DO NOT SUPPORT SPEEDS OVER 48000

0 8000 11025 12000 16000 22050 24000 32000 44100 48000 88200 96000 176400 192000 352800 384000

Proposal pitch threshold

Proposal pitch frequency threshold, for male models used at 155.0 and female models at 255.0

50 1200

Audio cleaning strength

Strength of the audio cleaner for filtering vocals during export

0 1

Autotune rate

Level of autotuning adjustment

0 1

Filter radius

If greater than three, median filtering is applied. The value represents the filter radius and can reduce breathiness or noise.

0 7

RMS Mix Rate

Determines the blend ratio between the RMS energy of the original voice and the converted voice

0 1

Protect

0 1

VAD sensitivity

VAD sensitivity. 0 is less sensitive, 3 is very sensitive.

0 3

VAD frame Size

VAD frame size processed for VAD (ms).

10 30

Frequency for Formant Shift 1

Frequency for Formant Shift

0 16

Timbre for Formant Transformation 1

Timbre for Formant Transformation

0 16

Frequency for Formant Shift 2

Frequency for Formant Shift

0 16

Timbre for Formant Transformation 2

Timbre for Formant Transformation

0 16

Audio input, output

Audio input

Conversion Output

Convert Text to Speech

Convert text to speech and read aloud using the trained voice model

Input data from a text file

Convert text using Google

Text To Speech

Voice speed

Speed of the voice

-100 100

Pitch

Recommendation: set to 12 to change male voice to female and vice versa

-24 24

Drop text file here

Voice Language

Pitch

Pitch adjustment for text-to-speech converter

-24 24

Input language

Output language

Model file

Index file

Index strength

Higher values increase strength. However, lower values may reduce artificial effects in the audio

0 1

Probe clusters

The number of clusters the algorithm will examine during approximate nearest neighbor search. Lower values are faster but less accurate, while higher values are more accurate but slower.

1 64

Speaker ID

Speaker ID for the multi-voice model

Export format

The export format to export the audio file in

wav mp3 flac ogg opus m4a mp4 aac alac wma aiff webm ac3

Output TTS path

Enter the output path

Converted TTS path

Enter the output path

Extracting pitch using the ONNX model can help improve speed

F0 ONNX

Unlock all pitch extraction methods

Unlock all

F0 method

Method used for data extraction

mangio-crepe-full crepe-full fcpe rmvpe harvest-stonemask pyin hybrid

HYBRID F0 method

Combination of 2 or more different types of extracts

Hop Length

Analyzing the time transfer window when performing transformations is allowed. The detailed value is compact but requires more calculation

64 512

Blending Level of Methods

The blending ratio of pitch values extracted by different methods.

0.1 1

Upload F0 file

F0 File

Blend multiple embedding layers together to achieve better audio quality.

Layer Mixing

Embedders Mode

Extracting embeddings using different models

fairseq onnx transformers

Embedding model

The extracts audio features (Note: Use only Contentvec or HuBERT Base; other models require separate training from scratch).

hubert_base contentvec_base vietnamese_hubert_base japanese_hubert_base korean_hubert_base chinese_hubert_base portuguese_hubert_base spin-v1 spin-v2 custom

Model name

If you have your own model, just upload it and input the name here

Layers to Mix

1 12

Mix Ratio

Controls how much each layer contributes to the final embedding. Higher values give more weight to higher layers.

0.1 1

Preset file

Save cleanup

Save autotune

Save pitch

Save index impact

Save resampling

Save median filter

Save sound envelope

Save sound protection

Save sound split

Pitch and Formant Shift

File name to save

Upload preset file

Input/output processing

Pitch and Formant Shift

Split audio

Clean audio

Autotune

Using memory-efficient training

Automatically propose pitch

Uses machine learning models to generate additional high-frequency spectrum layers, allowing the output audio to achieve a true sample rate instead of a virtual one created by resampling.

Audio Upscaler

Resample

Resample post-processing to the final sample rate; 0 means no resampling, NOTE: SOME FORMATS DO NOT SUPPORT SPEEDS OVER 48000

0 8000 11025 12000 16000 22050 24000 32000 44100 48000 88200 96000 176400 192000 352800 384000

Proposal pitch threshold

Proposal pitch frequency threshold, for male models used at 155.0 and female models at 255.0

50 1200

Autotune rate

Level of autotuning adjustment

0 1

Audio cleaning strength

Strength of the audio cleaner for filtering vocals during export

0 1

Filter radius

If greater than three, median filtering is applied. The value represents the filter radius and can reduce breathiness or noise.

0 7

RMS Mix Rate

Determines the blend ratio between the RMS energy of the original voice and the converted voice

0 1

Protect

0 1

Noise scale

A special adjustment that directly affects audio quality; you should not modify it.

0.1 1

Frequency for Formant Shift

0 16

Timbre for Formant Transformation

0 16

Unconverted and converted audio

TTS Output

Conversion Output

Add Additional Audio Effects

Add effects to audio

Reverb

Chorus

Delay

Phaser

Compressor

Additional options

Drop audio here

Audio input

Enter the path to the audio file

Audio output

Enter the output path

Merge instruments

Audio input

Enter the path to the audio file

Main audio file volume

Main audio file volume. Should be between -4 and 0.

-80 80

Combination audio file volume

Combination audio file volume. Should keep the volume of the combination file lower than the main audio.

-80 80

Export format

The export format to export the audio file in

wav mp3 flac ogg opus m4a mp4 aac alac wma aiff webm ac3

Preset file

File name to save

Upload preset file

Create a continuous echo effect when this mode is enabled

Freeze mode

Room size

Adjust the room space to create reverberation

0 1

Damping

Adjust the level of absorption to control the amount of reverberation

0 1

Reverb signal level

Adjust the level of the reverb signal effect

0 1

Original signal level

Adjust the level of the signal without effects

0 1

Audio width

Adjust the width of the audio space

0 1

Chorus depth

Adjust the intensity of the chorus to create a wider sound

0 1

Frequency

Adjust the oscillation speed of the chorus effect

0.1 10

Mix signals

Adjust the mix level between the original and the processed signal

0 1

Center delay (ms)

The delay time between stereo channels to create the chorus effect

0 50

Feedback

Adjust the amount of the effect signal fed back into the original signal

-1 1

Delay time

Adjust the delay time between the original and the processed signal

0 5

Delay feedback

Adjust the amount of feedback signal, creating a repeating effect

0 1

Delay signal mix

Adjust the mix level between the original and delayed signal

0 1

Phaser depth

Adjust the depth of the effect, impacting its intensity

0 1

Frequency

Adjust the frequency of the phaser effect

0.1 10

Mix signal

Adjust the mix level between the original and processed signals

0 1

Center frequency

The center frequency of the phaser effect, affecting the adjusted frequencies

50 5000

Feedback

Adjust the feedback level of the effect, creating a stronger or lighter phaser feel

-1 1

Compressor threshold

The threshold level above which the audio will be compressed

-60 0

Compression ratio

Adjust the level of audio compression when exceeding the threshold

1 20

Attack time (ms)

Time for compression to start taking effect after the audio exceeds the threshold

0.1 100

Release time

Time for the audio to return to normal after being compressed

10 1000

Fade

Bass and treble

Threshold limiter

Resample

Distortion

Audio gain

Bit reduction

Clipping

Fade-in (ms)

Time for the audio to gradually increase from 0 to normal level

0 10000

Fade-out (ms)

the time it takes for the sound to fade from normal to zero

0 10000

Bass boost (dB)

amount of bass boost in audio track

0 20

Low-pass filter cutoff frequency (Hz)

frequencies are reduced. Low frequencies make the bass clearer

20 200

Treble boost (dB)

high level of sound reinforcement in the audio track

0 20

High-pass filter cutoff frequency (Hz)

The frequency will be filtered out. The higher the frequency, the higher the sound will be retained.

1000 10000

Limiter threshold

Limit the maximum audio level to prevent it from exceeding the threshold

-60 0

Release time

Time for the audio to return after being limited (Mili Seconds)

10 1000

Pitch

Recommendation: set to 12 to change male voice to female and vice versa

-20 20

Resample

Resample post-processing to the final sample rate; 0 means no resampling, NOTE: SOME FORMATS DO NOT SUPPORT SPEEDS OVER 48000

0 8000 11025 12000 16000 22050 24000 32000 44100 48000 88200 96000 176400 192000 352800 384000

Distortion

Adjust the level of distortion to create a noisy effect

0 50

Audio gain

Adjust the volume level of the signal

-60 60

Clipping threshold

Trim signals exceeding the threshold, creating a distorted sound

-60 0

Bit depth

Reduce audio quality by decreasing bit depth, creating a distorted effect

1 24

Audio output

Audio input

Audio output

Weird Effects for Audio

Apply quirky effects to your audio to make it weird and weird.

Audio input

Quirk effects

Quirk effects that can be used to apply to audio

Random Voice Crack Horror Robot Baby Depression Voice Jerking Oldster Echo Devil Distorted Voice Online Sales Drag Uncomfortable Noise Connectivity Issue Disorder

Drop audio here

Export format

The export format to export the audio file in

wav mp3 flac ogg opus m4a mp4 aac alac wma aiff webm ac3

Input audio path

Enter the path to the audio file

Audio output path

Enter the output path (leave it as .wav format; it will auto-correct during conversion)

Audio output

Download Model

Download voice models, pretrained models

Choose a model download method

Download from url Download from CSV Search models Upload

Url to the model

Model name

Model repository

Name to search

Choose a searched model (Click to select)

Drop model here

Choose a model download method

Download from url Model list Upload

pretrained model link D

pretrained model link G

Choose pretrained model

Choose a pretrained model to download

Model samplerate

Drop pretrained model G, D here

Fushion 2 Models

Fushion 2 voice models into a single model

Model name

Model file 1

Model file 2

Model file 1

Model file 2

Model ratio

Adjusting towards one side will make the model more like that side

0 1

Model output path

Read Model Information

Retrieve recorded information within the model

Drop model here

Model file

Model Information

Converting PYTORCH Model to ONNX Model

Convert RVC model from pytorch to onnx to optimize audio conversion

Half-precision weight

Use Int8 for inference

Drop model here

Model file

Model output path

Convert SVC model to a readable project format

Convert SVC models trained from Sovits SVC 4.1 into a format readable by the project. Currently, only models with the original configuration are supported.

Delete original files when complete

Half-precision weight

Drop model here

Drop the configuration file here

Model file

Path to SVC configuration file

Model name

Model output path

Pitch Extraction

F0 pitch extraction is intended for use in audio conversion inference

Drop audio here

Audio input

F0 ONNX

Unlock all

Autotune

Automatically propose pitch

F0 method

Method used for data extraction

mangio-crepe-full crepe-full fcpe rmvpe harvest-stonemask pyin

Pitch

Recommendation: set to 12 to change male voice to female and vice versa

-24 24

Autotune rate

Level of autotuning adjustment

0 1

Proposal pitch threshold

Proposal pitch frequency threshold, for male models used at 155.0 and female models at 255.0

50 1200

Filter radius

If greater than three, median filtering is applied. The value represents the filter radius and can reduce breathiness or noise.

0 7

Input audio path

File

Image

Create SRT File From Audio File

Use Whisper to convert audio file to text and create srt file

SRT file content

Language

Set input languages to avoid the model confirming the wrong language

Whisper model size

Large models can produce strange outputs

tiny tiny.en base base.en small small.en medium medium.en large-v1 large-v2 large-v3 large-v3-turbo

Input audio path

Enter the path to the audio file

Output SRT file path

Drop audio here

Audio input

Output SRT file

Audio Information Analysis

Analyze information from an audio file

Input audio path

Enter the path to the audio file

Drop audio here

Audio input

Output Information

Image

Additional Settings

Customize additional features of the project

Language

The display language in the project (When making changes, a system restart is required for them to take effect)

Theme

Theme type displayed in the interface (When making changes, a system restart is required for them to take effect)

Font

Visit Google Font to choose font (When making changes, a system restart is required for them to take effect).

Compute device

Device used for computation and inference (System restart required to apply changes).

Use BF16 for training.

Use TF32 for training.

Use Int8 for inference

Precision

Precision of inference and model training

Note: CPU, OPENCL Does not support fp16

fp16 fp32

Commands

Use auxiliary commands to support development.

Model name

Name of the model during training (avoid special characters or spaces)

Click here if you want to be Rick Roll :) ---> RickRoll

Please do not use the project for any unethical, illegal, or harmful purposes to individuals or organizations...

In cases where users do not comply with the terms or violate them, I will not be responsible for any claims, damages, or liabilities, whether in contract, negligence, or other causes arising from, outside of, or related to the software, its use, or other transactions associated with it.

Built with Gradio logo