Fix PyTorch inference code: use libdf for correct feature extraction

README.md

@@ -31,18 +31,6 @@ pipeline_tag: audio-to-audio
 
 ---
 
-## Listen to the Model (Use headphones)
-
-**Raw Audio (Noisy Environment):**
-
-<audio controls src="https://huggingface.co/weya-ai/hush/resolve/main/assets/audio/sample_00006_raw.wav"></audio>
-
-**Denoised Audio (Hush Output):**
-
-<audio controls src="https://huggingface.co/weya-ai/hush/resolve/main/assets/audio/sample_00006_denoised.wav"></audio>
-
----
-
 ## Model Overview
 
 Hush is designed from the ground up for **Voice AI applications** — phone-based voice agents, call centre bots, voice assistants, real-time transcription pipelines, and conversational AI systems. It isolates exactly one speaker from a live audio stream, in real time, under production conditions.
@@ -146,25 +134,60 @@ ERB gain mask   Complex filter
 
 ## Quick Start: PyTorch Inference
 
+> **Important:** PyTorch inference requires `DeepFilterLib` for correct feature extraction.
+> Install it with `pip install DeepFilterLib`.
+
+The simplest way is the CLI script from the [GitHub repo](https://github.com/pulp-vision/Hush):
+
+```bash
+python scripts/infer_single.py \
+    --checkpoint model_best.ckpt \
+    --input noisy_speech.wav \
+    --output enhanced.wav
+```
+
+Or use the Python API directly:
+
 ```python
 import torch
+import numpy as np
 import soundfile as sf
-from model.dfnet_se import DfNetSE, get_config
+from libdf import DF, erb, erb_norm, unit_norm
+from model.dfnet_se import DfNetSE, as_complex, as_real, get_config, get_norm_alpha
 
+# Load model
 config = get_config()
 model = DfNetSE(config)
 checkpoint = torch.load("model_best.ckpt", map_location="cpu")
 model.model.load_state_dict(checkpoint)
 model.eval()
 
+# Load audio
 audio, sr = sf.read("noisy_speech.wav")
 assert sr == 16000, "Input must be 16 kHz"
-
-wav = torch.tensor(audio).float().unsqueeze(0).unsqueeze(0)  # [1, 1, T]
+wav = torch.tensor(audio, dtype=torch.float32).unsqueeze(0)  # [1, T]
+
+# Feature extraction via libdf (must match training pipeline)
+df_state = DF(sr=16000, fft_size=320, hop_size=160, nb_bands=32, min_nb_erb_freqs=2)
+alpha = get_norm_alpha(16000, 160, config.norm_tau)
+wav_padded = torch.nn.functional.pad(wav, (0, 320))
+spec_np = df_state.analysis(wav_padded.numpy(), reset=True)
+erb_feat = torch.as_tensor(erb_norm(erb(spec_np, df_state.erb_widths()), alpha)).unsqueeze(1)
+spec_feat = as_real(torch.as_tensor(unit_norm(spec_np[..., :64], alpha))).unsqueeze(1)
+spec_t = as_real(torch.as_tensor(spec_np)).unsqueeze(1)
+
+# Enhance
 with torch.no_grad():
-    enhanced = model(wav)  # [1, 1, T]
+    spec_enh = model.model(spec_t.clone(), erb_feat, spec_feat)[0]
+    spec_enh_c = as_complex(spec_enh.squeeze(1))
+
+# Synthesize and compensate delay
+enh_np = df_state.synthesis(spec_enh_c.numpy(), reset=True)
+enh = torch.from_numpy(np.asarray(enh_np, dtype=np.float32))
+delay = 320 - 160  # fft_size - hop_size
+enh = enh[:, delay : len(audio) + delay]
 
-sf.write("enhanced.wav", enhanced.squeeze().numpy(), 16000)
+sf.write("enhanced.wav", enh.squeeze().numpy(), 16000)
 ```
 
 ## Quick Start: Production (ONNX, No PyTorch)