Skip to content

flashsr.md

Latest commit

 

History

History
68 lines (50 loc) · 2.38 KB

flashsr.md

File metadata and controls

68 lines (50 loc) · 2.38 KB

FlashSR Inference

CLI command: speech upsample. Audio super-resolution at 48 kHz mono.

Quick start

# Build (release + metallib for best speed)
make build

# Upsample a low-bandwidth recording
.build/release/speech upsample noisy_lowres.wav -o clean_hr.wav

# Pick the int8 variant (slightly more accurate quantisation, ~720 MB on disk)
.build/release/speech upsample input.wav --variant int8 -o out.wav

# Reproducible noise (DPM-Solver init)
.build/release/speech upsample input.wav --seed 42 -o out.wav

Flags

--variant {int4,int8}   # default int4; both run at FP after load
--timestep N            # diffusion timestep (default 999 for the 1-step student)
--seed N                # random seed for the noise init
--output, -o PATH       # output WAV path (default hr.wav)

Input handling

  • Input can be any sample rate; it's resampled internally to 48 kHz mono.
  • Inputs are processed in non-overlapping 5.12 s windows.
  • Output length matches the input length (trimmed after windowed processing).
  • A single window is the most common case for music / SFX clips.

Programmatic use

import FlashSR

let model = try await FlashSR.fromPretrained(variant: .int4)
let hr = try model.upsample(audio: lowRes, sampleRate: 48_000)
try WAVWriter.write(samples: hr, sampleRate: 48_000,
                    to: URL(fileURLWithPath: "hr.wav"))

The model also conforms to SpeechEnhancementModel, so it slots into any enhance(audio:sampleRate:) consumer (e.g. an ASR pre-processor) — same signature as DeepFilterNet3 even though the semantics differ (super-resolution vs noise suppression). Pick by import: import FlashSR vs import SpeechEnhancement.

Pipeline

  1. Normalize: mean-center + max-abs scale to ±0.5.
  2. Log-mel spectrogram: HiFi-GAN-style (reflection pad, Hann window, 256 mels, Slaney filterbank, log(clamp(., 1e-5))).
  3. VAE encode: 8× downsample to a 16-channel latent, scaled by 0.3342.
  4. One-step DPM-Solver (v-prediction, cosine α̅): one UNet forward pass.
  5. VAE decode: mel back from the denoised latent.
  6. SR Vocoder: BigVGAN-style with SnakeBeta activations + audio conditioning pyramid; produces 48 kHz waveform.
  7. Denormalize: re-apply mean + scale.

License

Inherits FlashSR's upstream license: MIT (KAIST research). Output is free to use commercially; the weights themselves carry no use restrictions.