svc-develop-team/so-vits-svcArchived

Features

• VITS Synthesis Models - Implements a VITS-based generative architecture combining variational autoencoders and flow-based decoders for high-fidelity singing voice conversion.
• Voice Identity Conversions - Transforms singing audio into a target voice identity while preserving the original melody, lyrics, and intonation.
• Data Preparation - Resamples audio files, trims silence, and normalizes peak loudness to prepare vocal datasets for training.
• Feature Extractors - Predicts fundamental frequency and extracts content embeddings from audio to facilitate voice conversion.
• Voice Model Trainers - Provides a framework for preprocessing audio datasets and training VITS models to capture specific vocal characteristics.
• Voice Model Merging - Blends multiple voice models or speaker identities to create unique hybrid vocal identities through linear combinations.
• Voice Synthesizer Training - Processes audio datasets and trains deep learning models to replicate specific vocal characteristics and timbres.
• Pitch-Guided Synthesis - Uses fundamental frequency predictors to map source pitch and maintain melodic accuracy during voice conversion.
• Training Data Generation - Converts raw audio files into spectrograms, speaker IDs, and text sequences for deep learning model training.
• Vocal Content Embeddings - Extracts speaker-independent latent representations from raw audio to separate linguistic content from vocal identity.
• Vocal Timbre Mixers - Blends multiple speaker models to create hybrid voice identities through linear interpolation.
• Audio Feature Extraction - Extracts content embeddings, fundamental frequency, and volume from audio files to prepare for voice conversion.
• Pitch Estimation - Employs predictor algorithms to estimate the pitch of an audio signal for use in voice conversion.
• AI Audio Enhancement - Applies diffusion probabilistic models as post-processors to remove electronic artifacts and improve vocal clarity.
• Dataset Scanning Tools - Scans audio directories to generate file lists while filtering short clips and mapping speaker identities.
• Vocal Denoising Post-processors - Applies a diffusion probabilistic model as a post-processor to remove electronic artifacts and improve vocal clarity.
• Voice Index Generators - Extracts speaker-specific features from audio datasets to create index files that guide the conversion process.
• Voice Identity Interpolators - Blends multiple speaker identities via weighted averages of model weights or feature embeddings for hybrid voice creation.
• Hybrid Voice Synthesis - Mixes multiple target speaker characteristics to create a hybrid vocal identity for output audio.
• Voice Model Loaders - Enables the import of external model and configuration files to define specific target vocal characteristics.
• Acoustic Feature Retrieval - Uses cluster-based nearest-neighbor retrieval of acoustic features to improve target voice similarity and reduce timbre leakage.
• Time-Stretching and Pitch-Shifting - Shifts the pitch of converted audio by specific semitones to modify the musical key of the output.
• Vocal Pitch Normalization - Adjusts the fundamental frequency of voice recordings to a standard scale for consistent voice conversion.
• Timbre Fidelity Controllers - Balances retrieval-based index files to reduce timbre leakage and increase similarity to the target voice.
• Vocal Artifact Removal - Applies diffusion models and enhancers to reduce electronic artifacts and improve the clarity of converted vocals.
• Voice Processing - Framework for singing voice conversion using deep learning.

Star history