Audio generation research has a reproducibility problem: text-to-speech, singing synthesis, and voice conversion each grew their own ecosystem of single-purpose repos, incompatible data pipelines, and bespoke evaluation scripts. Amphion's bet is that these tasks share far more than they differ — the same vocoders, the same feature extractors, the same metrics — so it puts them under one roof and turns any-input-to-audio into a single, comparable workflow.
What Sets It Apart
- Breadth under one framework: TTS, singing voice synthesis (SVS), voice conversion (VC), singing voice conversion (SVC), accent conversion, and text-to-audio all share infrastructure, so swapping a vocoder or comparing two acoustic models doesn't mean rebuilding the stack.
- Classic-model visualizations: rather than only shipping weights, it diagrams how landmark architectures actually work — a deliberate on-ramp for junior researchers who would otherwise reverse-engineer papers.
- A unified evaluation suite: objective metrics for speech and singing live next to the models, so results are reproducible against a shared yardstick instead of each paper's private script.
Great Fit / Look Elsewhere
Great fit if you're a researcher or grad student who wants reproducible baselines across several audio-generation tasks, or you want to compare vocoders and acoustic models on equal footing. Look elsewhere if you need a turnkey production TTS API — Amphion is a research toolkit, recipe- and training-heavy, and assumes comfort with PyTorch, GPUs, and reading papers rather than calling a hosted endpoint.
