FlashSR Go

Upstream / Original Project: ysharma3501/FlashSR

This Go implementation is based on and inspired by the original FlashSR project.

flashsr-go is a Go library and CLI for FlashSR audio super-resolution. It runs the FlashSR ONNX model to convert speech-like audio to 48 kHz output.

• Library API for in-process audio pipelines.
• CLI for WAV input/output workflows.
• Streaming wrapper with overlap/crossfade behavior.
• Optional resampling for non-16 kHz inputs.
• Browser demo kernel compiled to Go/WASM.

Requirements

• Go 1.25+
• ONNX Runtime shared library (libonnxruntime)
  • Linux: .so
  • macOS: .dylib
  • Windows: .dll

FlashSR uses github.com/yalue/onnxruntime_go, which loads ONNX Runtime dynamically. Set the library path with FLASHSR_ORT_LIB or --ort-lib.

Quick Start (CLI)

Build:

go build -o flashsr ./cmd/flashsr

Set ONNX Runtime path:

export FLASHSR_ORT_LIB=/path/to/libonnxruntime.so

Run environment checks:

./flashsr doctor

Upsample a WAV file:

./flashsr upsample --input in.wav --output out_48k.wav

Use explicit input rate override:

./flashsr upsample \
  --input in.wav \
  --output out_48k.wav \
  --input-rate 24000

Use streaming mode:

./flashsr upsample \
  --input in.wav \
  --output out_48k.wav \
  --stream \
  --chunk-size 4000

CLI Commands

• flashsr upsample: WAV -> 48 kHz WAV
• flashsr doctor: model + ONNX Runtime diagnostics
• flashsr model download: fetch model from Hugging Face

Download model manually:

./flashsr model download --out assets/model.onnx

Useful environment variables:

• FLASHSR_ORT_LIB: path to libonnxruntime
• FLASHSR_MODEL_PATH: override embedded model file
• HF_TOKEN: Hugging Face token for gated/private model access

Library Usage

package main

import (
	"os"

	"github.com/MeKo-Christian/flashsr-go/flashsr"
)

func main() {
	u, err := flashsr.New(flashsr.Config{
		ORTLibPath: os.Getenv("FLASHSR_ORT_LIB"),
		InputRate:  24000, // auto-resample to 16 kHz before inference
	})
	if err != nil {
		panic(err)
	}
	defer u.Close()

	in := []float32{0, 0.1, -0.1} // example PCM
	out, err := u.Upsample16kTo48k(in)
	if err != nil {
		panic(err)
	}

	_ = out
}

Web Demo (Go/WASM)

The repository includes a browser demo in web/ with an explicit wasm kernel in cmd/flashsr-wasm.

Behavior:

• Upload a WAV file.
• Analyze sample rate quality.
• If input sample rate is below 48 kHz, resample to 48 kHz.
• If input sample rate is >= 48 kHz, keep the original WAV bytes.

Local run:

./web/build-wasm.sh
python3 -m http.server 8080 -d web/dist

or:

just web-demo

Open: http://localhost:8080

GitHub Pages deploy is defined in:

• .github/workflows/deploy-pages.yml

It builds web/dist/flashsr-kernel.wasm, copies wasm_exec.js, and uploads web/dist as the Pages artifact.

Development

Run tests:

go test ./...

Useful just targets:

• just test
• just lint
• just ci
• just web-demo

Notes

• Embedded model: assets/model.onnx
• Pinned model hash is model.ExpectedSHA256 in model/model.go.
• The ORT backend currently reports provider CPU.

License

MIT license for this repository code.

See:

• LICENSE
• NOTICE
• THIRD_PARTY_NOTICES.md