upload LiSenNet checkpoint + ONNX (fp32, int8 dyn/static)

.gitattributes

@@ -33,3 +33,4 @@ 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
+g_best filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,102 @@
 ---
 license: mit
+library_name: pytorch
+tags:
+- speech-enhancement
+- audio
+- denoising
+- onnx
+- causal
+- streaming
+- real-time
+datasets:
+- JacobLinCool/VoiceBank-DEMAND-16k
 ---
+
+# LiSenNet
+
+An ultra-compact (~37 K-param), causal, real-time speech enhancer trained on
+VoiceBank-DEMAND-16k. Sub-band U-Net with a dual-path-recurrent bottleneck and a
+magnitude-only mask; phase is recovered with a 2-iteration Griffin-Lim (offline)
+or reused from the noisy phase (real-time).
+
+Faithful port of **Yan, Zhou, Chen & Lu, _LiSenNet: Lightweight Sub-band and
+Dual-Path Modeling for Real-Time Speech Enhancement_,
+[arXiv:2409.13285](https://arxiv.org/abs/2409.13285)** (authors' MIT reference:
+[hyyan2k/LiSenNet](https://github.com/hyyan2k/LiSenNet)).
+
+Training / export / quantization code and the full write-up:
+[https://github.com/LarocheC/eco8-neaixt](https://github.com/LarocheC/eco8-neaixt) — see
+[RESULTS_LISENNET.md](https://github.com/LarocheC/eco8-neaixt/blob/main/RESULTS_LISENNET.md).
+
+## Headline numbers
+
+PESQ is wideband, on the full 824-utterance VoiceBank-DEMAND test split.
+
+| metric                                  |     value |
+| --------------------------------------- | --------: |
+| params                                  | 36,783    |
+| FP32 PESQ (torch / ONNX, Griffin-Lim)   | **3.006** |
+| dynamic-int8 PESQ (Griffin-Lim)         |     2.995 |
+| **real-time PESQ (int8 + noisy phase)** | **2.982** |
+| static-int8 PESQ (Griffin-Lim)          |     2.897 |
+| RTF (1 thread CPU)                      |     0.15  |
+
+The ONNX export is loss-free; dynamic weight-only int8 costs ~0.012 PESQ. The
+reproduction lands within ~0.06 of the paper's reported ~3.07.
+
+## Files
+
+| file                      | what it is |
+| ------------------------- | --- |
+| `g_best`                  | PyTorch checkpoint (`{"generator": state_dict}`) |
+| `g_best_fp32.onnx`        | FP32 ONNX of the mask sub-network: `feat (B,3,T,F)` -> `est_mag (B,T,F)`, dynamic batch/time |
+| `g_best_int8_dyn.onnx`    | Dynamic weight-only int8 (robust, near loss-free) — **recommended** |
+| `g_best_int8_static.onnx` | Static full int8 (QDQ, per-tensor, percentile calibration; QAT-grade) |
+| `config.json`             | Training + architecture config (STFT, channels, blocks) |
+
+The ONNX graphs cover only the mask sub-network (the pure tensor-op core). The
+STFT, the 3-channel feature extraction
+(`[compressed_mag, group_delay/pi, instantaneous_freq_diff/pi]`), the magnitude
+combination, and the phase recovery (Griffin-Lim offline / noisy phase for
+real-time) live outside the graph — see the source repo's
+`lisennet/streaming.py` and `lisennet/eval_deploy.py`.
+
+## Loading
+
+PyTorch:
+
+```python
+import json, torch
+from huggingface_hub import hf_hub_download
+from common.env import AttrDict
+from lisennet.model import build_lisennet
+
+REPO = "claroche1/LiSenNet"
+cfg  = json.load(open(hf_hub_download(REPO, "config.json")))
+ckpt = torch.load(hf_hub_download(REPO, "g_best"),
+                  map_location="cpu", weights_only=True)
+model = build_lisennet(AttrDict(cfg)).eval()
+model.load_state_dict(ckpt["generator"])
+# end-to-end waveform -> waveform: model(noisy_wav)["est"]
+```
+
+ONNX (FP32 or int8) — the mask sub-network:
+
+```python
+import numpy as np, onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+REPO = "claroche1/LiSenNet"
+sess = ort.InferenceSession(
+    hf_hub_download(REPO, "g_best_int8_dyn.onnx"),   # or _fp32 / _int8_static
+    providers=["CPUExecutionProvider"],
+)
+# input  feat: (B, 3, T, F=257) = [compressed_mag, group_delay/pi, ifd/pi]
+# output est_mag: (B, T, F)  -> pair with the noisy phase + iSTFT for audio
+est_mag = sess.run(["est_mag"], {"feat": np.zeros((1, 3, 100, 257), np.float32)})[0]
+```
+
+## License
+
+MIT. See the [source repository](https://github.com/LarocheC/eco8-neaixt) for training code and full attribution.

config.json

@@ -0,0 +1,36 @@
+{
+    "_comment": "LiSenNet (arXiv:2409.13285, hyyan2k/LiSenNet, MIT) ported into this framework. Hyperparameters follow the upstream config.yaml: VoiceBank-DEMAND, n_fft=512/hop=256, 2 s segments, AdamW lr 5e-4 betas [0.8,0.99], StepLR gamma 0.98/epoch, CMGAN losses (complex 0.1, mag 0.9, adv 0.05). num_channels=16, n_blocks=2 -> ~37k-param generator. Phase via Griffin-Lim (no phase loss). batch_size is 16 (not the paper's 4): the model is tiny (0.77 GB at batch 16) but each step pays a fixed CPU-PESQ + Griffin-Lim cost, so batch 4 is needlessly slow (~32 min/epoch vs ~4.3 at batch 16); per-sample norms make MetricGAN batch-robust.",
+
+    "num_gpus": 0,
+    "batch_size": 16,
+    "learning_rate": 0.0005,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.98,
+    "seed": 1234,
+
+    "sampling_rate": 16000,
+    "segment_size": 32000,
+
+    "n_fft": 512,
+    "hop_size": 256,
+    "win_size": 512,
+
+    "num_channels": 16,
+    "n_blocks": 2,
+    "compress_factor": 0.3,
+    "gradient_clip": 5.0,
+
+    "loss": {
+        "complex": 0.1,
+        "mag": 0.9,
+        "adv": 0.05
+    },
+
+    "num_workers": 4,
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

g_best

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f983e79d7018ec7e0df6d5cb457b2178d78c019c759f266cc5bb83f463e0b8d5
+size 177557

g_best_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b381ddff78913380499711a49e2b6b5b2bb31129bcc65c9971b09119197a13a
+size 286688

g_best_int8_dyn.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cfdb82a597c5bcee3466dc731b6f2e5260237255d95e16d80bcdd117ced305b
+size 330599

g_best_int8_static.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a2e23114faa0bc56afce246a2b46b684d08967553f19a431f2a139235bb5b21
+size 453595