FlashSR
This first-party Soniqo page documents FlashSR from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.
Internal Page First
Landing cards and docs menus now point here first; source model and bundle links remain available from this page.
At a Glance
| Model | FlashSR |
|---|---|
| Role | Audio super-resolution for low-bandwidth or lossy audio |
| Backend | MLX int4 default; int8 available |
| Output | 48 kHz mono waveform, same length as input |
| Languages | Audio-content agnostic |
| License | MIT |
| Status | Ready through speech upsample and the FlashSR Swift product |
| Source | FlashSR / AudioSR distillation |
| Swift product | FlashSR |
| CLI / runtime | speech upsample |
Use
The snippet below mirrors the current speech-swift API or command exposed by the repo.
# Upsample a low-bandwidth recording to 48 kHz mono.
.build/release/speech upsample noisy_lowres.wav \
--variant int4 \
-o clean_hr.wavModel Links
Implementation Notes
- Download requests the single model.safetensors bundle and config explicitly; this is simple enough that byte-weighting is less critical than for sharded models.
- Input is resampled to 48 kHz mono and processed in non-overlapping 5.12 second windows.
- INT4 is a download-size optimization; runtime weights dequantize to FP, so memory footprint matches int8.
- The model conforms to SpeechEnhancementModel, but its semantics are super-resolution rather than denoising.