Dataset-first workflow for roughly 30 minutes of raw vocal material.
Neural Singing Voice Platform
A modular audio ML platform for singing-voice conversion, covering dataset validation, model adapters, durable jobs, evaluation, an API, and a local interface.
Designed the platform around reproducible datasets, isolated model adapters, portable compute backends, and a clear separation between implemented capabilities and optional model-dependent workflows.
01 / PROBLEM
What the system is trying to solve.
Singing-voice workflows often become tightly coupled to one model, one GPU backend, or manually prepared audio. The platform treats data preparation, training, inference, evaluation, and device selection as separate reusable layers.
02 / SYSTEM
How it works.
- 01
Prepare owned vocal material and validate datasets before training.
- 02
Keep model-specific logic behind adapters instead of spreading it through the application.
- 03
Run durable training and inference jobs through a backend service with explicit state.
- 04
Detect available compute and support portable backend selection across CUDA and AMD-capable environments.
- 05
Expose the workflow through an API and local product interface rather than notebook-only experiments.
03 / ENGINEERING
Decisions that shaped the project.
Model adapter architecture for swapping singing-voice or voice-conversion backends.
Portable compute strategy across AMD and NVIDIA systems.
Separation and resynthesis workflow for owned music material.
Evaluation and job-state boundaries designed for reproducible experimentation.
04 / EVIDENCE
A few concrete signals.
Singing voice conversion
Primary domain~30 min raw vocals
Training inputAMD + NVIDIA aware
Compute strategyAPI + local UI
Interface05 / LINKS