Add deepfilternet

docs/deepfilternet.md

@@ -0,0 +1,61 @@
+# DeepFilterNet Input Filter
+
+The engine can optionally run DeepFilterNet on inbound microphone audio before
+Pipecat VAD and STT. The integration uses DeepFilterNet's real-time `libDF` C
+API (`df_process_frame`) rather than the Python `df.enhance()` batch helper.
+
+## Build DeepFilterNet
+
+From the DeepFilterNet checkout:
+
+```bash
+cd /Users/wangx/Code/DeepFilterNet
+cargo build --release -p deep_filter --features capi
+```
+
+Use the generated native library path as `audio_filter.lib_path`. On macOS this
+is usually:
+
+```text
+/Users/wangx/Code/DeepFilterNet/target/release/libdf.dylib
+```
+
+Use an ONNX tar.gz model as `audio_filter.model_path`, for example:
+
+```text
+/Users/wangx/Code/DeepFilterNet/models/DeepFilterNet3_ll_onnx.tar.gz
+```
+
+The low-latency model is preferred for a live voice endpoint.
+
+## Install Optional Python Dependencies
+
+```bash
+uv pip install -r requirements-deepfilternet.txt
+```
+
+## Enable
+
+```json
+"audio_filter": {
+  "enabled": true,
+  "provider": "deepfilternet",
+  "lib_path": "/Users/wangx/Code/DeepFilterNet/target/release/libdf.dylib",
+  "model_path": "/Users/wangx/Code/DeepFilterNet/models/DeepFilterNet3_ll_onnx.tar.gz",
+  "model_sample_rate_hz": 48000,
+  "atten_lim_db": 100.0,
+  "post_filter_beta": 0.0,
+  "log_level": null
+}
+```
+
+`model_sample_rate_hz` defaults to `48000`, matching the bundled DeepFilterNet
+models. The filter resamples from the engine sample rate to the model sample
+rate, processes hop-sized frames, then resamples back to the engine sample rate.
+
+You can also provide paths through environment variables:
+
+```bash
+export DEEPFILTERNET_LIB_PATH=/Users/wangx/Code/DeepFilterNet/target/release/libdf.dylib
+export DEEPFILTERNET_MODEL_PATH=/Users/wangx/Code/DeepFilterNet/models/DeepFilterNet3_ll_onnx.tar.gz
+```