UniverSR.AXERA AX650 8k/16k speech SR: 6 axmodels, C++ bin, Python SDK

.gitattributes

@@ -33,3 +33,14 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+models/universr_split_16k_condenc.axmodel filter=lfs diff=lfs merge=lfs -text
+models/universr_split_16k_main_cond_actual.axmodel filter=lfs diff=lfs merge=lfs -text
+models/universr_split_16k_main_uncond.axmodel filter=lfs diff=lfs merge=lfs -text
+models/universr_split_8k_condenc.axmodel filter=lfs diff=lfs merge=lfs -text
+models/universr_split_8k_main_cond_actual.axmodel filter=lfs diff=lfs merge=lfs -text
+models/universr_split_8k_main_uncond.axmodel filter=lfs diff=lfs merge=lfs -text
+samples/listen/OSR_us_000_0010_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/16k_output_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/8k_output_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/speech_2_input_16k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/speech_2_output_48k.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,88 @@
+# UniverSR.AXERA
+
+**8 / 16 kHz speech → 48 kHz · AX650 NPU3 · C++ RTF 4.38 / Python RTF 4.3 · 6 个 axmodel 共 381 MB**
+
+UniverSR 是 vocoder-free flow matching 超分模型，本包包含 **8k 和 16k** 两档语音模型。
+
+- GitHub 源码：https://github.com/ZY-2012/UniverSR.AXERA
+- HuggingFace：https://huggingface.co/HY-2012/UniverSR.AXERA
+- 上游：https://github.com/woongzip1/UniverSR
+
+## 目录
+
+- 快速开始
+- 结果
+- 推理速度
+- 数据来源
+- License
+
+## 快速开始（AX650 板端）
+
+```bash
+# C++（默认）
+bash run_ax650.sh samples/listen/OSR_us_000_0010_input_8k_5s.wav output_8k.wav --sr 8
+bash run_ax650.sh samples/listen/speech_2_input_16k.wav output_16k.wav --sr 16
+
+# Python SDK
+bash run_ax650.sh samples/listen/OSR_us_000_0010_input_8k_5s.wav output_8k.wav --sr 8 --python
+```
+
+输入采样率必须和 `--sr` 一致；输出固定 48 kHz。默认 STFT 128 帧分块、overlap 64 帧，Flow ODE 4 步，guidance 1.5。
+
+## 结果
+
+### English 8 kHz
+
+| 输入 | UniverSR 输出 |
+|---|---|
+| <audio controls style="width:220px" src="https://huggingface.co/HY-2012/UniverSR.AXERA/resolve/main/samples/listen/OSR_us_000_0010_input_8k_5s.wav"></audio><br><img src="samples/listen/spectrograms/8k_input_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/UniverSR.AXERA/resolve/main/samples/listen/OSR_us_000_0010_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/8k_output_spec.png" width="360"> |
+
+### English 16 kHz
+
+| 输入 | UniverSR 输出 |
+|---|---|
+| <audio controls style="width:220px" src="https://huggingface.co/HY-2012/UniverSR.AXERA/resolve/main/samples/listen/speech_2_input_16k.wav"></audio><br><img src="samples/listen/spectrograms/16k_input_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/UniverSR.AXERA/resolve/main/samples/listen/speech_2_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/16k_output_spec.png" width="360"> |
+
+## 推理速度
+
+| 运行时 | 5.12 s 输入 | RTF |
+|---|---:|---:|
+| **C++** | **22–25 s** | **4.3–4.9** |
+| Python SDK | 22 s | 4.3 |
+
+## 精度
+
+| 输入 | 板端 vs PyTorch |
+|---|---:|
+| 8k → 48k | **0.99193** |
+| 16k → 48k | **0.99350** |
+
+## 工程结构
+
+```text
+UniverSR.AXERA/
+├── bin/                 # C++ 板端可执行文件
+├── models/              # 6 个 axmodel + model_meta.json
+├── python/              # Python SDK
+├── samples/             # 中英文输入 / 输出 / 频谱图
+├── run_ax650.sh
+├── configuration.json
+├── config.json          # {}
+└── requirements.txt
+```
+
+## 说明
+
+- 8k/16k 使用不同 `lr_bins`（80/170），模型不能混用。
+- `main_cond` 使用真实 `cond encoder` 输出做校准，`vc` 才能达到 0.9999。
+- C++ 和 Python 都按 128 帧分块、overlap 64 帧。
+- 当前版本只提供 8k/16k，12k/24k 未包含。
+
+## 数据来源
+
+- 上游 UniverSR：https://github.com/woongzip1/UniverSR
+- 示例语音：Open Speech Repository、公开 16 kHz 人声样例。
+
+## License
+
+MIT；上游 UniverSR 权重遵循其原始许可证。

config.json

@@ -0,0 +1 @@
+{}

configuration.json

@@ -0,0 +1,13 @@
+{
+  "framework": "axera",
+  "model": "UniverSR-speech",
+  "target_hardware": "AX650",
+  "npu_mode": "NPU3",
+  "compiler": "Pulsar2 8.0",
+  "task": "audio-to-audio",
+  "sample_rate_out": 48000,
+  "input_sample_rates": [8000, 16000],
+  "models": ["universr_split_8k_condenc.axmodel", "universr_split_8k_main_cond_actual.axmodel", "universr_split_8k_main_uncond.axmodel", "universr_split_16k_condenc.axmodel", "universr_split_16k_main_cond_actual.axmodel", "universr_split_16k_main_uncond.axmodel"],
+  "python_demo": "python/demo.py",
+  "cpp_bin": "bin/universr_ax650"
+}

models/model_meta.json

@@ -0,0 +1,12 @@
+{
+  "name": "UniverSR-speech",
+  "input_sr": [8000, 16000],
+  "output_sr": 48000,
+  "stft_frames": 128,
+  "overlap_frames": 64,
+  "n_fft": 1024,
+  "hop_length": 512,
+  "ode_steps": 4,
+  "guidance_scale": 1.5,
+  "compiler": "Pulsar2 8.0"
+}

models/universr_split_16k_condenc.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ade59b27cbc99500d9b6c9815af246d9fe2a13599cd01b720e5836859ab629a
+size 17072502

models/universr_split_16k_main_cond_actual.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d42eeb782be3ce535db7be58c38b2c92b81e7ed67b080bc179b5dd74d682dca4
+size 71718411

models/universr_split_16k_main_uncond.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1baa0d79e20a642e4d7f710932029d65d1c7ad1fd36f4df73131133507247b63
+size 113505290

models/universr_split_8k_condenc.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01805afed4e514f9d5b46ddf9b316e6089eb54d518314a5f22b3ccabe2293511
+size 11059108

models/universr_split_8k_main_cond_actual.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:957b904bb939287a709ad8552ce8bbc281b09a9ba20c2af80bb83436e7d0c80f
+size 71718993

models/universr_split_8k_main_uncond.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2cbda2d420f0bcb8ca70ac63c5dfc0ba958ccd908dc642a278ba4121f790d1
+size 113504544

python/demo.py

@@ -0,0 +1,48 @@
+import sys,time,argparse
+from pathlib import Path
+import numpy as np, torch, torchaudio, soundfile as sf, axengine
+ROOT=Path(__file__).resolve().parent; sys.path.insert(0,str(ROOT))
+T=128; F_HR=432; HOP=512; NFFT=1024; SR=48000
+WIN=torch.hann_window(NFFT)
+def stft(x):
+ X=torch.stft(x.reshape(1,-1),NFFT,HOP,window=WIN,center=True,return_complex=True)
+ A=X+1e-4; X=A.abs()**0.2*torch.exp(1j*A.angle())
+ r=torch.view_as_real(X).permute(0,3,1,2); return r[:,:,:-1,:]
+def istft(sp):
+ sp=torch.nn.functional.pad(sp,[0,0,0,1],value=0).permute(0,2,3,1).contiguous()
+ X=torch.view_as_complex(sp); X=X.abs()**5.0*torch.exp(1j*X.angle())
+ return torch.istft(X,NFFT,HOP,window=WIN,center=True,length=245760)
+def main():
+ p=argparse.ArgumentParser(); p.add_argument('input'); p.add_argument('output'); p.add_argument('--models',default=str(ROOT/'models')); p.add_argument('--seed',type=int,default=42); p.add_argument('--overlap',type=int,default=64); p.add_argument('--sr',type=int,default=8); a=p.parse_args()
+ ce=axengine.InferenceSession(str(Path(a.models)/f'universr_split_{a.sr}k_condenc.axmodel'))
+ mc=axengine.InferenceSession(str(Path(a.models)/f'universr_split_{a.sr}k_main_cond_actual.axmodel'))
+ mu=axengine.InferenceSession(str(Path(a.models)/f'universr_split_{a.sr}k_main_uncond.axmodel'))
+ x,sr=sf.read(a.input); x=torch.tensor(x,dtype=torch.float32).reshape(1,-1)
+ x=torchaudio.functional.resample(x,sr,SR); x=torch.nn.functional.pad(x,(0,max(0,245760-x.shape[-1])))[...,:245760]
+ Y=stft(x.unsqueeze(1)); total=Y.shape[-1]; lr=80 if a.sr==8 else 170; hf=512-432
+ pad=(T-total%T)%T
+ Yp=torch.nn.functional.pad(Y,(0,pad),mode='constant') if pad else Y
+ total=Yp.shape[-1]
+ starts=list(range(0,total-T+1,T-a.overlap))
+ spec=torch.zeros_like(Yp); wsum=torch.zeros(total)
+ rng=np.random.RandomState(a.seed); t0=time.time()
+ for s in starts:
+  y=Yp[:,:,:lr,s:s+T].numpy().astype(np.float32)
+  z=ce.run(None,{'y':y})[0].astype(np.float32)  # 1,384,1,T
+  zz=rng.randn(1,2,F_HR,T).astype(np.float32)
+  n=128
+  for i in range(4):
+   t=np.array([i/4],np.float32); dt=0.25
+   def f(q,tt):
+    vu=mu.run(None,{'x':q,'t':tt})[0]; vc=mc.run(None,{'x':q,'t':tt,'y_cond':z})[0]
+    return (1-1.5)*vu+1.5*vc
+   k1=f(zz,t); zm=zz+0.5*dt*k1; k2=f(zm,np.array([i/4+0.125],np.float32)); zz=zz+dt*k2
+  cut=max(0,lr-hf); full=np.concatenate([y,zz[:,:,cut:,:]],axis=2) # 1,2,512,T?
+  full=torch.from_numpy(full)
+  w=torch.ones(T)
+  if s>0: w[:a.overlap]=torch.linspace(0,1,a.overlap)
+  if s+T>total: w[-(s+T-total):]=0
+  spec[:,:,:,s:s+T]+=full*w; wsum[s:s+T]+=w
+ spec=spec/wsum.clamp_min(1e-8); out=istft(spec).squeeze()[:245760]
+ sf.write(a.output,out.numpy(),SR,subtype='PCM_16'); print('time',time.time()-t0,'shape',out.shape)
+if __name__=='__main__': main()

python/requirements.txt

@@ -0,0 +1,5 @@
+numpy
+soundfile
+torch
+torchaudio
+pyaxengine

requirements.txt

@@ -0,0 +1,5 @@
+numpy
+soundfile
+torch
+torchaudio
+pyaxengine

run_ax650.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+set -e
+cd "$(dirname "$0")"
+export LD_LIBRARY_PATH=/soc/lib:${LD_LIBRARY_PATH:-}
+INPUT="${1:-samples/listen/OSR_us_000_0010_input_8k_5s.wav}"
+OUTPUT="${2:-output.wav}"
+if [ $# -ge 2 ]; then shift 2; else shift $#; fi
+SR=8; MODE=cpp; EXTRA=()
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --sr) SR="$2"; shift 2 ;;
+    --python) MODE=python; shift ;;
+    --overlap) EXTRA+=(--overlap "$2"); shift 2 ;;
+    *) EXTRA+=("$1"); shift ;;
+  esac
+done
+if [ "$MODE" = cpp ] && [ -x ./bin/universr_ax650 ]; then
+  exec ./bin/universr_ax650 models "$INPUT" "$OUTPUT" --sr "$SR"
+fi
+exec python3 python/demo.py "$INPUT" -o "$OUTPUT" --models models --sr "$SR" --overlap 64 "${EXTRA[@]}"

samples/listen/OSR_us_000_0010_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a26e83365a420bda90e5a9ef3cbf35e6766bdf1d814232d240650756fcf1499
+size 491564

samples/listen/speech_2_input_16k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c7773c82ce75aa5e8a14d5bb53ee5915dbb65f47272ae9f527ac767c7b1f80c
+size 163884

samples/listen/speech_2_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc747b84d164e0c1d9b48a3cee7fe277ff5480fea5fbead30068980a79f6db05
+size 491564