Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
65 lines
2.4 KiB
Markdown
65 lines
2.4 KiB
Markdown
# AudioSR
|
|
|
|
AudioSR performs audio super-resolution from an input waveform.
|
|
|
|
## Quick Start
|
|
|
|
```bash
|
|
audiocpp_cli \
|
|
--task s2s \
|
|
--family audiosr \
|
|
--model models/AudioSR-GGUF/audiosr-basic-f32.gguf \
|
|
--backend cuda \
|
|
--audio input.wav \
|
|
--request-option num_inference_steps=50 \
|
|
--request-option guidance_scale=3.5 \
|
|
--request-option ddim_eta=1.0 \
|
|
--request-option seed=42 \
|
|
--out output.wav
|
|
```
|
|
|
|
## Model
|
|
|
|
| Field | Value |
|
|
|---|---|
|
|
| Family | `audiosr` |
|
|
| Task | `s2s` |
|
|
| Mode | `offline` |
|
|
| Default package | `models/AudioSR-GGUF/audiosr-basic-f32.gguf` |
|
|
| Input | Source WAV through `--audio` |
|
|
| Output | Super-resolved waveform |
|
|
|
|
`ddim_eta` defaults to `1.0`, matching the official AudioSR inference default.
|
|
Use `--request-option ddim_eta=0` only for deterministic sampler debugging.
|
|
|
|
Long audio is processed with bounded overlapping chunks once the input exceeds
|
|
`audio_chunk_duration_sec`. The defaults match the official Python long-audio
|
|
path:
|
|
|
|
```bash
|
|
--request-option audio_chunk_duration_sec=15 \
|
|
--request-option audio_chunk_overlap_sec=2
|
|
```
|
|
|
|
## Common Options (use directly)
|
|
|
|
| Option | Values | Default | Meaning |
|
|
|---|---|---:|---|
|
|
| `--audio` | WAV path | required | Source waveform to super-resolve. |
|
|
|
|
## Request Options (use with `--request-option`)
|
|
|
|
| Option | Values | Default | Meaning |
|
|
|---|---|---:|---|
|
|
| `num_inference_steps` | integer > 0 | `50` | DDIM inference steps. The uniform DDIM schedule must produce exactly this many timesteps; not every count is valid (e.g. 51 fails with 1000 diffusion timesteps). |
|
|
| `guidance_scale` | float | `3.5` | Classifier-free guidance scale. Any finite value is accepted. |
|
|
| `ddim_eta` | float >= 0 | `1.0` | DDIM eta; 1.0 matches the official AudioSR default. Must be nonnegative; there is no runtime upper bound. |
|
|
| `audio_chunk_duration_sec` | seconds > 0 | `15` | Long-audio chunk size; must exceed the overlap. |
|
|
| `audio_chunk_overlap_sec` | seconds >= 0 | `2` | Overlap must be smaller than `audio_chunk_duration_sec`, including after rounding to samples. |
|
|
| `seed` | integer <= 2147483647 | `42` | Sampler seed. Omitted or negative values use fixed seed 42, not a random seed; zero is a fixed seed. |
|
|
|
|
## Session Options (use with `--session-option`)
|
|
|
|
| Option | Values | Default | Meaning |
|
|
|---|---|---:|---|
|
|
| `audiosr.weight_type` | `native`, `f32`, `f16`, `bf16`, `q8_0` | `native` | Weight storage override for experiments. |
|