FlowHigh AX650 U16/S8 quantized models, Python runtime and samples

.gitattributes

@@ -33,3 +33,57 @@ 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/bigvgan.axmodel filter=lfs diff=lfs merge=lfs -text
+models/flowhigh.axmodel filter=lfs diff=lfs merge=lfs -text
+samples/listen/16k/speech_2_input_16k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/16k/speech_3_input_16k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from16k/speech_2_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from16k/speech_3_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_cn_000_0072_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_cn_000_0073_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_cn_000_0074_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_cn_000_0075_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_us_000_0010_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_us_000_0013_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_us_000_0057_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/OSR_us_000_0059_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/samplelib_speech_1m_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/speech_2_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/48k_from8k/speech_3_output_48k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_cn_000_0072_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_cn_000_0073_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_cn_000_0074_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_cn_000_0075_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_us_000_0010_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_us_000_0013_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_us_000_0057_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/OSR_us_000_0059_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/8k/samplelib_speech_1m_input_8k.wav filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0072_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0072_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0073_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0073_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0074_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0074_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0075_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_cn_000_0075_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0010_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0010_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0013_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0013_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0057_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0057_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0059_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/OSR_us_000_0059_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/samplelib_speech_1m_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/samplelib_speech_1m_output_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_2_input_16k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_2_input_8k.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_2_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_2_output_16k_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_2_output_8k_48k.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_2_output_8k_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_3_input_16k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_3_input_8k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_3_output_16k_48k_spec.png filter=lfs diff=lfs merge=lfs -text
+samples/listen/spectrograms/speech_3_output_8k_48k_spec.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Jun-Hak Yun
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,145 @@
+---
+library_name: axera
+pipeline_tag: audio-to-audio
+base_model: kiriyamaX/FLowHigh_indep_adaptive_400k
+language:
+- en
+- zh
+tags:
+- axera
+- ax650
+- npu3
+- flowhigh
+- speech-super-resolution
+- audio-super-resolution
+- bandwidth-extension
+- real-time
+- python
+---
+
+# FlowHigh.AXERA
+
+**8/16 kHz 语音 → 48 kHz 超分 · AX650 NPU3 · Python RTF 0.24（20–60 s 音频）· 2 个 axmodel 共 58 MB · 不做降噪**
+
+![RTF](https://img.shields.io/badge/RTF-0.24%20Python-brightgreen) ![Size](https://img.shields.io/badge/axmodel-58%20MB-green) [![GitHub](https://img.shields.io/badge/GitHub-ZY--2012%2FFlowHigh.AXERA-black)](https://github.com/ZY-2012/FlowHigh.AXERA)
+
+- GitHub: https://github.com/ZY-2012/FlowHigh.AXERA
+- 上游: https://github.com/jjunak-yun/FLowHigh_code
+
+## 目录
+
+- 快速开始
+- 结果
+- 推理速度
+- 数据来源
+- License
+
+## 快速开始（AX650 板端）
+
+从 HuggingFace 下载 `models/`、`assets/`、`python/`、`samples/`，在包根目录运行：
+
+```bash
+bash run_ax650.sh samples/listen/8k/OSR_us_000_0010_input_8k.wav output.wav
+```
+
+输入支持 8 kHz / 16 kHz，输出固定 48 kHz。默认 1-step flow、`std=1e-4`；不做降噪。
+
+## 结果
+
+### 8 kHz 输入（Open Speech / Samplelib）
+
+#### English
+
+**OSR_us_000_0010**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_us_000_0010_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0010_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_us_000_0010_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0010_output_48k_spec.png" width="360"> |
+
+**OSR_us_000_0013**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_us_000_0013_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0013_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_us_000_0013_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0013_output_48k_spec.png" width="360"> |
+
+**OSR_us_000_0057**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_us_000_0057_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0057_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_us_000_0057_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0057_output_48k_spec.png" width="360"> |
+
+**OSR_us_000_0059**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_us_000_0059_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0059_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_us_000_0059_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_us_000_0059_output_48k_spec.png" width="360"> |
+
+#### Chinese
+
+**OSR_cn_000_0072**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_cn_000_0072_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0072_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_cn_000_0072_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0072_output_48k_spec.png" width="360"> |
+
+**OSR_cn_000_0073**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_cn_000_0073_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0073_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_cn_000_0073_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0073_output_48k_spec.png" width="360"> |
+
+**OSR_cn_000_0074**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_cn_000_0074_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0074_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_cn_000_0074_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0074_output_48k_spec.png" width="360"> |
+
+**OSR_cn_000_0075**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/OSR_cn_000_0075_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0075_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/OSR_cn_000_0075_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/OSR_cn_000_0075_output_48k_spec.png" width="360"> |
+
+#### Samplelib（1 分钟，频谱图为前 10 s）
+
+**samplelib_speech_1m**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/samplelib_speech_1m_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/samplelib_speech_1m_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/samplelib_speech_1m_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/samplelib_speech_1m_output_48k_spec.png" width="360"> |
+
+### 16 kHz 参照（人声）
+
+同一批人声降采样到 16 kHz 后输入；8k 行为 8 kHz 输入，16k 行为 16 kHz 输入。
+
+**speech_2**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/speech_2_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/speech_2_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/speech_2_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/speech_2_output_8k_48k_spec.png" width="360"> |
+| 16k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/16k/speech_2_input_16k.wav"></audio><br><img src="samples/listen/spectrograms/speech_2_input_16k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from16k/speech_2_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/speech_2_output_16k_48k_spec.png" width="360"> |
+
+**speech_3**
+
+| 版本 | 输入 | FlowHigh 输出 |
+|---|---|---|
+| 8k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/8k/speech_3_input_8k.wav"></audio><br><img src="samples/listen/spectrograms/speech_3_input_8k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from8k/speech_3_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/speech_3_output_8k_48k_spec.png" width="360"> |
+| 16k | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/16k/speech_3_input_16k.wav"></audio><br><img src="samples/listen/spectrograms/speech_3_input_16k_spec.png" width="360"> | <audio controls style="width:220px" src="https://huggingface.co/HY-2012/FlowHigh.AXERA/resolve/main/samples/listen/48k_from16k/speech_3_output_48k.wav"></audio><br><img src="samples/listen/spectrograms/speech_3_output_16k_48k_spec.png" width="360"> |
+
+## 推理速度：RTF
+
+| 链路 | Python RTF |
+|---|---:|
+| 8/16 kHz → 48 kHz | **0.24** |
+
+20–60 s 音频，1-step flow，BigVGAN 1 s 分块、0.2 s overlap。5 s 短音频因分块/启动开销会略高；C++ 运行时尚未提供。
+
+## 数据来源
+
+- Open Speech Repository（English / Chinese 8k）：http://voiptroubleshooter.com/open_speech/index.html
+- Samplelib speech 1m：https://samplelib.com/mp3/sample-speech-1m.mp3
+- 16 kHz 参照：公开人声样例经 48k->16k 降采样。
+
+## License
+
+MIT；FLowHigh 与 BigVGAN 均为 MIT。模型 checkpoint 来自上游/公开镜像，仓库不重复分发训练权重。

assets/time_emb_4.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06284a683d5610206adac80e4764a61f1d77d8cd4d172376513ad5c83d965f0f
+size 16512

config.json

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

models/bigvgan.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fb5f3a130ef732eea8ef9ac9f365e7bab95d34d176f5926f4a3d21db33306b5
+size 20734384

models/flowhigh.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ed959fdc9632a961204eb3ea8872395102505db96d42730a8c0e1fb6189bd12
+size 38297185

python/demo_board.py

@@ -0,0 +1,107 @@
+import argparse
+import time
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+
+from flowhigh_sdk import audio
+
+
+def decode_mel(voc, mel, chunk, overlap):
+    hop = chunk - overlap
+    wav = None
+    for s in range(0, mel.shape[-1], hop):
+        m = mel[..., s:s + chunk]
+        if m.shape[-1] < chunk:
+            m = np.pad(m, ((0, 0), (0, 0), (0, chunk - m.shape[-1])), mode="edge")
+        c = voc.run(None, {"mel": np.ascontiguousarray(m)})[0][0, 0]
+        if wav is None:
+            wav = c
+        else:
+            n = min(overlap * 480, len(wav), len(c))
+            if n > 0:
+                fade = np.linspace(0.0, 1.0, n, dtype=np.float32)
+                wav[-n:] = wav[-n:] * (1.0 - fade) + c[:n] * fade
+            wav = np.concatenate([wav, c[n:]])
+    return wav.astype(np.float32)
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--input", required=True)
+    p.add_argument("--output", required=True)
+    p.add_argument("--models", default="models")
+    p.add_argument("--steps", type=int, default=1)
+    p.add_argument("--std", type=float, default=1e-4)
+    p.add_argument("--seed", type=int, default=0)
+    p.add_argument("--chunk", type=int, default=100)
+    p.add_argument("--overlap", type=int, default=20)
+    p.add_argument("--flow-overlap", type=int, default=50)
+    p.add_argument("--time-emb", default="assets/time_emb_4.npy")
+    args = p.parse_args()
+
+    x, sr = sf.read(args.input, always_2d=False)
+    if x.ndim > 1:
+        x = x.mean(axis=1)
+    x = np.asarray(x, dtype=np.float32)
+    target = int(round(audio.SR / sr * x.size))
+    x = audio.resample(x, sr)
+    peak = float(np.max(np.abs(x)))
+    if peak > 0:
+        x = x / peak
+    cond_wave = x.astype(np.float32)
+    mel = audio.log_mel(cond_wave)[0].T[None]
+    frames = 500
+    if mel.shape[-1] < frames:
+        mel = np.pad(mel, ((0, 0), (0, 0), (0, frames - mel.shape[-1])), mode="edge")
+    total = mel.shape[-1]
+    if total > frames:
+        n_chunks = (total - frames + (frames - args.flow_overlap) - 1) // (frames - args.flow_overlap) + 1
+        padded = frames + (n_chunks - 1) * (frames - args.flow_overlap)
+        if padded > total:
+            mel = np.pad(mel, ((0, 0), (0, 0), (0, padded - total)), mode="edge")
+    padded_total = mel.shape[-1]
+
+    import axengine
+
+    flow = axengine.InferenceSession(str(Path(args.models) / "flowhigh.axmodel"))
+    voc = axengine.InferenceSession(str(Path(args.models) / "bigvgan.axmodel"))
+    emb = np.load(args.time_emb).astype(np.float32)
+    rng = np.random.RandomState(args.seed)
+    hop = frames - args.flow_overlap
+    starts = list(range(0, max(1, padded_total - frames + 1), hop))
+
+    t0 = time.time()
+    flow_ms = []
+    out = None
+    for s in starts:
+        c = mel[..., s:s + frames]
+        y = c + args.std * rng.randn(*c.shape).astype(np.float32)
+        for i in range(args.steps):
+            t1 = time.time()
+            v = flow.run(None, {
+                "x": np.ascontiguousarray(y),
+                "cond": np.ascontiguousarray(c),
+                "time_emb": np.ascontiguousarray(emb[i]),
+            })[0]
+            flow_ms.append((time.time() - t1) * 1000)
+            y = y + (1.0 / args.steps) * v
+        wav = decode_mel(voc, y, args.chunk, args.overlap)
+        if out is None:
+            out = wav
+        else:
+            n = min(args.flow_overlap * 480, len(out), len(wav))
+            if n > 0:
+                fade = np.linspace(0.0, 1.0, n, dtype=np.float32)
+                out[-n:] = out[-n:] * (1.0 - fade) + wav[:n] * fade
+            out = np.concatenate([out, wav[n:]])
+    out = out[:int(total * 480)]
+    out = audio.postprocess(out, cond_wave[:len(out)], len(out))[:target]
+    sf.write(args.output, out, audio.SR, subtype="PCM_16")
+    dt = time.time() - t0
+    print(f"chunks={len(starts)} flow_ms={flow_ms} total_ms={dt*1000:.1f} rtf={dt/(x.size/audio.SR):.4f}")
+
+
+if __name__ == "__main__":
+    main()

python/flowhigh_sdk/audio.py

@@ -0,0 +1,138 @@
+import numpy as np
+from math import gcd
+
+SR = 48000
+N_FFT = 2048
+HOP = 480
+N_MELS = 256
+FMIN = 20.0
+FMAX = 24000.0
+LOG_FLOOR = 1e-5
+MAG_EPS = 1e-9
+
+_MEL = None
+_WIN = None
+
+
+def _hz_to_mel(f):
+    f = np.asarray(f, dtype=np.float64)
+    f_sp, min_log_hz = 200.0 / 3.0, 1000.0
+    min_log_mel = min_log_hz / f_sp
+    logstep = np.log(6.4) / 27.0
+    mel = np.empty_like(f)
+    lin = f < min_log_hz
+    mel[lin] = f[lin] / f_sp
+    mel[~lin] = min_log_mel + np.log(f[~lin] / min_log_hz) / logstep
+    return mel
+
+
+def _mel_to_hz(m):
+    m = np.asarray(m, dtype=np.float64)
+    f_sp, min_log_hz = 200.0 / 3.0, 1000.0
+    min_log_mel = min_log_hz / f_sp
+    logstep = np.log(6.4) / 27.0
+    hz = np.empty_like(m)
+    lin = m < min_log_mel
+    hz[lin] = m[lin] * f_sp
+    hz[~lin] = min_log_hz * np.exp(logstep * (m[~lin] - min_log_mel))
+    return hz
+
+
+def mel_filterbank():
+    global _MEL
+    if _MEL is None:
+        freqs = np.linspace(0.0, SR / 2.0, N_FFT // 2 + 1)
+        edges = np.linspace(_hz_to_mel([FMIN])[0], _hz_to_mel([FMAX])[0], N_MELS + 2)
+        hz = _mel_to_hz(edges)
+        fb = np.zeros((N_MELS, freqs.shape[0]), dtype=np.float32)
+        for m in range(N_MELS):
+            left, center, right = hz[m], hz[m + 1], hz[m + 2]
+            up = (freqs - left) / (center - left)
+            down = (right - freqs) / (right - center)
+            fb[m] = np.maximum(0.0, np.minimum(up, down)) * (2.0 / max(1e-8, right - left))
+        _MEL = fb
+    return _MEL
+
+
+def hann():
+    global _WIN
+    if _WIN is None:
+        _WIN = np.hanning(N_FFT + 1)[:-1].astype(np.float64)
+    return _WIN
+
+
+def log_mel(audio):
+    x = np.asarray(audio, dtype=np.float64).reshape(-1)
+    pad = (N_FFT - HOP) // 2
+    x = np.pad(x, (pad, pad), mode="reflect")
+    frames = 1 + (x.size - N_FFT) // HOP
+    idx = np.arange(N_FFT)[None, :] + HOP * np.arange(frames)[:, None]
+    spec = np.fft.rfft(x[idx] * hann(), n=N_FFT, axis=1).T
+    mag = np.sqrt(spec.real ** 2 + spec.imag ** 2 + MAG_EPS)
+    mel = np.log(np.clip(mel_filterbank() @ mag, LOG_FLOOR, None))
+    return mel.T[None].astype(np.float32)
+
+
+def stft(x, center=True, pad_mode="constant"):
+    x = np.asarray(x, dtype=np.float64).reshape(-1)
+    if center:
+        p = N_FFT // 2
+        x = np.pad(x, (p, p), mode=pad_mode)
+    frames = 1 + (x.size - N_FFT) // HOP
+    idx = np.arange(N_FFT)[None, :] + HOP * np.arange(frames)[:, None]
+    return np.fft.rfft(x[idx] * hann(), n=N_FFT, axis=1).T
+
+
+def istft(spec, length=None, center=True):
+    spec = np.asarray(spec)
+    frames = np.fft.irfft(spec.T, n=N_FFT, axis=1) * hann()
+    out_len = N_FFT + HOP * (frames.shape[0] - 1)
+    ola = np.zeros(out_len, dtype=np.float64)
+    wsum = np.zeros(out_len, dtype=np.float64)
+    w2 = hann() ** 2
+    for i in range(frames.shape[0]):
+        s = i * HOP
+        ola[s:s + N_FFT] += frames[i]
+        wsum[s:s + N_FFT] += w2
+    ola /= np.where(wsum > 1e-11, wsum, 1.0)
+    if center:
+        ola = ola[N_FFT // 2:len(ola) - N_FFT // 2]
+    if length is not None:
+        if ola.size > length:
+            ola = ola[:length]
+        elif ola.size < length:
+            ola = np.pad(ola, (0, length - ola.size))
+    return ola.astype(np.float32)
+
+
+def cutoff_bin(spec):
+    energy = np.cumsum(np.abs(spec).sum(axis=-1))
+    if energy[-1] <= 0:
+        return 0
+    limit = energy[-1] * 0.99
+    below = np.nonzero(energy < limit)[0]
+    return int(below[-1]) + 1 if below.size else 0
+
+
+def postprocess(pred, src, length):
+    ps = stft(pred, center=True, pad_mode="constant")
+    ss = stft(src, center=True, pad_mode="constant")
+    n = min(ps.shape[-1], ss.shape[-1])
+    ps, ss = ps[:, :n], ss[:, :n]
+    cr = cutoff_bin(ss)
+    merged = ps.copy()
+    merged[:cr] = ss[:cr]
+    out = istft(merged, length=length, center=True)
+    peak = np.max(np.abs(out))
+    if peak > 0:
+        out = out / peak * 0.99
+    return out.astype(np.float32)
+
+
+def resample(x, sr_in, sr_out=SR):
+    x = np.asarray(x, dtype=np.float32).reshape(-1)
+    if sr_in == sr_out:
+        return x
+    g = gcd(int(sr_out), int(sr_in))
+    from scipy.signal import resample_poly
+    return resample_poly(x, sr_out // g, sr_in // g).astype(np.float32)

python/requirements.txt

@@ -0,0 +1,3 @@
+numpy
+scipy
+soundfile

reports/quantization.md

@@ -0,0 +1,67 @@
+# FlowHigh AX650 quantization report
+
+## Target
+
+- Board: AX650N, NPU3, axengine 2.12.0s, compiler 8.0.
+- Flow: U16 activations / S8 weights, 38 MB.
+- BigVGAN: U16 activations / S8 weights, 20 MB.
+- Input: 8 kHz / 16 kHz -> 48 kHz.
+- No denoising.
+
+## Graph split
+
+```text
+host: resample_poly -> log-mel [1,256,T] -> 1-step flow -> BigVGAN chunks -> STFT merge
+NPU:  flow velocity (37 MB) + BigVGAN 1 s chunks (20 MB)
+```
+
+The flow NPU input is channel-first `[1,256,500]`; the wrapper transposes to `[1,500,256]` internally. This is required: Pulsar2 lowers 3D tensors channel-first, and feeding `[1,500,256]` directly makes attention run on the wrong axis.
+
+## Numerical checks
+
+| Check | Result |
+|---|---|
+| Flow ONNX vs PyTorch velocity | max err 1.5e-4 |
+| Flow U16/S8 on board vs ONNX | SNR 32.96 dB, corr 0.99977 |
+| Flow latency, 5 s | 21 ms (after warmup) |
+| BigVGAN ONNX vs PyTorch | max err 8.2e-6 |
+| BigVGAN U16/S8 on board vs ONNX, 1 s | SNR 20.55 dB, corr 0.99560 |
+| BigVGAN latency, 1 s | ~93 ms |
+| BigVGAN chunk overlap | 0.2 s; boundary jump reduced from 0.44 to 0.005 |
+
+BigVGAN is the sensitive part; its direct waveform SNR is low, but the final speech quality after spectral merging is close to the FP32 pipeline.
+
+## Full pipeline quality
+
+VCTK `p232_013`, 48 kHz reference. Board uses 1-step flow with `std=1e-4`, BigVGAN 1 s / 0.2 s overlap. Baseline and board use different RNG seeds, so the PESQ/LSD comparison is indicative; flow and BigVGAN SNR rows are strict same-input comparisons.
+
+| Input | Pipeline | PESQ-WB | STOI | Mel LSD |
+|---|---|---|---:|---:|
+| 8 kHz | PyTorch FP32 | 3.872 | 0.9981 | 2.323 |
+| 8 kHz | AX650 U16/S8 | **3.876** | **0.9982** | 2.333 |
+| 16 kHz | PyTorch FP32 | 4.570 | 1.0000 | 2.315 |
+| 16 kHz | AX650 U16/S8 | **4.573** | **1.0000** | 2.302 |
+
+## Speed
+
+5 s input, Python/axengine, one flow step:
+
+| Stage | Time |
+|---|---:|
+| Flow | 33 ms |
+| BigVGAN (6 x 1 s chunks) | 578 ms |
+| Host STFT/mel/postprocess | ~600 ms |
+| Total | 0.94–1.22 s |
+| RTF | 0.19–0.24 |
+
+The current runtime is Python-only. Moving STFT/postprocess to C++ and using a native flow/vocoder loop is the next optimization; the NPU portion itself is 0.61 s for 5 s.
+
+## Long audio
+
+The flow graph is fixed at 500 mel frames (5 s). Longer audio is processed in 5 s chunks with 0.5 s overlap and waveform crossfade; BigVGAN uses 1 s chunks with 0.2 s overlap. The 60 s Samplelib input was processed end-to-end with no boundary clicks above the normal waveform derivative.
+
+## Limitations
+
+- The 5 s flow chunk is a hard graph size; longer context would need a different compiled graph.
+- No denoising; FLowHigh is BWE only.
+- C++ runtime is not included in this first quantization milestone.

run_ax650.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+set -e
+cd "$(dirname "$0")"
+if [ "$#" -ge 2 ] && [ "${1#-}" = "$1" ]; then
+  IN="$1"; OUT="$2"; shift 2
+  exec python3 python/demo_board.py --models models --steps 1 --std 0.0001 --input "$IN" --output "$OUT" "$@"
+fi
+exec python3 python/demo_board.py --models models --steps 1 --std 0.0001 "$@"

samples/listen/16k/speech_2_input_16k.wav

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

samples/listen/16k/speech_3_input_16k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19cb7158760a383a60f1c2e5de23a780331f492fd5a7ccc63a58cadb8ad2f804
+size 163884

samples/listen/48k_from16k/speech_2_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dddb1324a6bd655a1d3e3eebe426ebe7fd1f27cecb5691a8a1c9c65a6900b78
+size 491564

samples/listen/48k_from16k/speech_3_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea134e4f9a4c1ce8c683f1b4c8d867614c48fd9c7d86012c5b28f4eac929eba9
+size 491564

samples/listen/48k_from8k/OSR_cn_000_0072_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3aacb425e8d754df060a22409847c89bae4100837eae32326e7d206bcc607afa
+size 1916204

samples/listen/48k_from8k/OSR_cn_000_0073_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ccd4b0a34f5a6cc92c944410cb2d20a68ccf4a875f3a34ef8597ac2a205b638
+size 2102444

samples/listen/48k_from8k/OSR_cn_000_0074_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f34689207317fa1b6d6bbf5cbf9f8214fca395f7883161dea7258e4e3b4a91a7
+size 2207084

samples/listen/48k_from8k/OSR_cn_000_0075_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e5845b5def78204cfae9d3fb4dfa5bf8604bf2f0537ad5abdf2cb7c57b77822
+size 2389484

samples/listen/48k_from8k/OSR_us_000_0010_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:020b7219f36a61dc11ea0b003a8249ee7de2b8ed4cb4b0a14c27f7250836c18a
+size 3227564

samples/listen/48k_from8k/OSR_us_000_0013_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d23124c3e9a8415a9ca48de6b3c432fb2534f91d660b0ef680a86fdc156aa2c4
+size 3123884

samples/listen/48k_from8k/OSR_us_000_0057_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11c50ba9a97219286e7e15aaf58043083e871fa36eafdeb0cc571656f821a23a
+size 3768044

samples/listen/48k_from8k/OSR_us_000_0059_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b72a16bda85904b21786ecea508bbe847e1579dc2c34f558284ff765bbd36a2
+size 5348204

samples/listen/48k_from8k/samplelib_speech_1m_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f87bcb7e77799e390f1017231d35931bef3c6d7c1262334c7cc775189a21bdbe
+size 5760044

samples/listen/48k_from8k/speech_2_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba2dcda972762f4369839be6776fd77c7291edfd00a7b164ef8ab433963e3641
+size 491564

samples/listen/48k_from8k/speech_3_output_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6aa796c46e7f7938c242018ba818c97dabedf29cf74c02c298be73658d6ff77b
+size 491564

samples/listen/8k/OSR_cn_000_0072_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7e2e6c7a0476d0286c6b1df92d15ab938888dfecd73efb29e4a0b82740fa761
+size 319526

samples/listen/8k/OSR_cn_000_0073_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffacc5400d699ff1b84917308a4f9f0eb996a7957d4ee4bb126c7d3a6ee1e375
+size 350544

samples/listen/8k/OSR_cn_000_0074_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef91f9f846a94d5e617fdba65a9acbb1eaffcbbafee50880e92bb857e171c7f
+size 367988

samples/listen/8k/OSR_cn_000_0075_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fd3b56518d3aed721b5e3c98c7492796eb85d225b52ee325053d9d595edcb5e
+size 398442

samples/listen/8k/OSR_us_000_0010_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4bf9becd046d7aedb6d05b6e12347a6294a44f74d263089c636fb0a2b1e6561
+size 538014

samples/listen/8k/OSR_us_000_0013_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3af3f5df07da603b7a421a345aad023df28c862ba92f68af360cc8c64c71436e
+size 520730

samples/listen/8k/OSR_us_000_0057_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebf286a093ca75e7f53b26180c6e54e7f179680b19d57a52f19516dd036381fe
+size 628192

samples/listen/8k/OSR_us_000_0059_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84520e52d7f8e7fadbb0d3bfbf18ddf80ff8bd7862d9287a48bfe8060223e554
+size 891442

samples/listen/8k/samplelib_speech_1m_input_8k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41fa574399ab8fa2f9263be1d386449f884627d205e72c310233c481cacf9b57
+size 960164