| |
| """FastWave.AXERA demo: any sample rate -> 48 kHz super resolution on AX650 NPU.""" |
| import argparse |
| import os |
| import sys |
| import time |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| sys.path.insert(0, HERE) |
|
|
| from fastwave_sdk import FastWaveSR |
| from fastwave_sdk import audio as A |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser(description="FastWave.AXERA 48 kHz audio super resolution") |
| ap.add_argument("input", help="input wav (6 kHz - 48 kHz)") |
| ap.add_argument("output", help="output wav (48 kHz)") |
| ap.add_argument("--model", default=os.path.join(HERE, "..", "models", "fastwave_nfe.axmodel")) |
| ap.add_argument("--chunk", type=int, default=None, help="chunk samples at 48 kHz (must match the axmodel, default from model_config.json)") |
| ap.add_argument("--overlap", type=float, default=0.5, help="chunk overlap ratio") |
| ap.add_argument("--steps", type=int, default=8, help="EDM sampling steps (4 or 8)") |
| ap.add_argument("--seed", type=int, default=1234) |
| ap.add_argument("--denoise", action="store_true", help="enable the spectral-subtraction noise post filter") |
| args = ap.parse_args() |
|
|
| wav, sr = A.read_wav(args.input) |
| print(f"[input ] {args.input} {len(wav)/sr:.2f}s @ {sr} Hz") |
| model = FastWaveSR(args.model, num_steps=args.steps, chunk=args.chunk, |
| overlap=args.overlap, seed=args.seed) |
| t0 = time.time() |
| out = model.enhance(wav, sr, denoise=args.denoise, |
| progress=lambda i, n: print(f"\r[chunk ] {i}/{n}", end="", flush=True)) |
| dt = time.time() - t0 |
| print() |
| A.write_wav(args.output, out, 48000) |
| dur = len(out) / 48000.0 |
| print(f"[output] {args.output} {dur:.2f}s @ 48000 Hz") |
| print(f"[perf ] {dt:.2f}s RTF {dt / dur:.3f}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
