livekit-plugins-hush 0.3.3__tar.gz

livekit_plugins_hush-0.3.3/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

livekit_plugins_hush-0.3.3/PKG-INFO

@@ -0,0 +1,196 @@
+Metadata-Version: 2.4
+Name: livekit-plugins-hush
+Version: 0.3.3
+Summary: Hush noise suppression plugin for LiveKit Agents
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/arpesenti/livekit-plugins-hush
+Project-URL: Repository, https://github.com/arpesenti/livekit-plugins-hush
+Project-URL: Bug Tracker, https://github.com/arpesenti/livekit-plugins-hush/issues
+Keywords: livekit,noise-suppression,noise-cancellation,hush,speech,audio,voip
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: livekit>=1.0.25
+Requires-Dist: livekit-agents>=1.4.4
+Requires-Dist: numpy>=1.26.0
+Requires-Dist: onnxruntime>=1.17.0
+Dynamic: license-file
+
+# livekit-plugins-hush
+
+[LiveKit](https://livekit.io) noise suppression plugin using the [Hush](https://github.com/pulp-vision/Hush) speech enhancement model. Self-hosted, in-process, no cloud API.
+
+Hush is built on [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet) with an auxiliary separation head for background speaker suppression. Inference uses a pure-numpy DSP frontend for STFT/ERB feature extraction and [ONNX Runtime](https://onnxruntime.ai) for the neural network. No PyTorch, no Rust toolchain, no prebuilt mystery binaries.
+
+---
+
+## Installation
+
+```
+pip install livekit-plugins-hush
+```
+
+Dependencies: `livekit >= 1.0.25`, `livekit-agents >= 1.4.4`, `numpy >= 1.26.0`, `onnxruntime >= 1.17.0`.
+
+---
+
+## Usage
+
+```python
+from livekit.agents import room_io
+from livekit.plugins import hush
+
+await session.start(
+    room_options=room_io.RoomOptions(
+        audio_input=room_io.AudioInputOptions(
+            noise_cancellation=hush.noise_suppression(),
+        ),
+    ),
+)
+```
+
+One instance per session. The ONNX model is loaded once per worker process and shared across instances.
+
+### Parameters
+
+```python
+hush.noise_suppression(
+    strength=0.5,       # wet/dry blend: 0.0 = bypass, 1.0 = full (default 0.5)
+    atten_lim_db=100.0, # max attenuation in dB (100.0 = unlimited)
+    debug_logging=False, # log per-frame RMS every 100 frames
+)
+```
+
+---
+
+## Architecture
+
+The model operates at 16 kHz with 10 ms frames (160 samples, 320-sample FFT). Processing is **per-frame streaming**: one 160-sample frame in, one 160-sample frame out, matching the API shape of the upstream `weya_nc` C library.
+
+The three ONNX sub-models (`enc`, `erb_dec`, `df_dec`) have been re-exported to expose the encoder, ERB decoder, and DF decoder GRU hidden states as I/O. `HushSession` holds those three states (and a 4-frame DF filter history) and threads them through every call, so the GRU has continuous memory across the entire session — same mechanism as the C library.
+
+The pure-numpy STFT runs in streaming mode (`reset=False`), so its analysis/synthesis filter state is carried across frames. The first frame's output is the STFT warmup (effectively zero); algorithmic latency is **1 frame (10 ms)** for the STFT lookahead, on top of GRU-state propagation which is fully causal.
+
+### Signal flow
+
+```
+LiveKit AudioFrame (any rate, any channels)
+  → resample to 16 kHz mono
+  → buffer 10 ms (160 samples) per frame
+  → Pure-numpy streaming STFT (1 frame out, state carried)
+  → ERB features + DF spectral features
+  → ONNX: encoder → ERB decoder + DF decoder (with continuous GRU state)
+  → apply ERB mask + DF complex filter to spectrum
+  → apply atten_lim_db linear blend (matches upstream reference)
+  → Pure-numpy streaming ISTFT
+  → wet/dry blend
+  → upsample, restore channels
+→ AudioFrame
+```
+
+### Model sharing
+
+The ONNX session is a process-level singleton. Each `HushNoiseSuppressor` instance holds only per-stream DF state and audio buffers.
+
+| Per worker | Per session |
+|---|---|
+| ~40 MB (onnxruntime + model) | ~1 MB (DF state + buffers) |
+
+---
+
+## Inference performance
+
+| | Per-frame streaming | Batch (full file) |
+|---|---|---|
+| Frame size | 1 frame (10 ms) | Full audio |
+| Algorithmic latency | 10 ms | N/A (offline) |
+| Per-frame time (single session) | ~0.25 ms | — |
+| Per-frame time (20 concurrent) | ~0.20 ms | — |
+| Real-time factor | 0.025× (per frame) | 0.008× |
+| Throughput | ~4,000 frames/sec per core (single session) | 129× real-time |
+| Concurrent sessions per core | ~390+ at 100 fps | — |
+| Model size | ~9 MB (3 ONNX files) | |
+
+Per-frame time measured on a single ARM64 core with the silero-style
+low-latency ORT config (`intra_op_num_threads=1`, `inter_op_num_threads=1`,
+`ORT_SEQUENTIAL`, no spinning waits). End-to-end `process_frame` takes
+~0.35 ms per 10 ms frame (~28× real-time headroom on a single core).
+Of that, ~0.085 ms is the pure-numpy DSP (STFT/ISTFT + ERB projection
++ per-band EMA) and ~0.27 ms is the three ONNX sub-model inferences
+(encoder + ERB decoder + DF decoder). For typical LiveKit use cases
+(5-20 concurrent calls per agent worker), the plugin is comfortably
+real-time on a single core.
+
+**Concurrency model:** the plugin is GIL-bound. N concurrent sessions
+on one core share the same Python GIL; per-frame work is mostly
+C-level (ORT) which releases the GIL, plus a small amount of pure-numpy
+DSP (STFT + ERB projection + per-band EMA). For higher concurrency
+than ~500 streams per core, run multiple agent worker processes.
+
+**ORT config:** we use the same low-latency session options as the
+upstream [silero VAD plugin](https://github.com/livekit/agents/tree/main/livekit-plugins/livekit-plugins-silero)
+— `intra_op_num_threads=1`, `inter_op_num_threads=1`,
+`execution_mode=ORT_SEQUENTIAL`, no spinning waits. This is ~2× faster
+than ORT defaults for single-stream inference because it avoids the
+per-op thread-pool overhead ORT enables by default.
+
+---
+
+## ONNX model
+
+The ONNX model bundle is from the public PyTorch checkpoint ([weya-ai/hush](https://huggingface.co/weya-ai/hush)). The three sub-models are re-exported with GRU hidden state as I/O using `scripts/export_onnx_stateful.py`; the export script downloads the PyTorch weights from Hugging Face and exports the new ONNX bundle.
+
+Output parity is verified via `scripts/verify_against_pytorch.py`, which compares the per-frame ONNX pipeline output against the original PyTorch model.
+
+## Benchmarking
+
+`scripts/benchmark.py` measures per-frame latency, multi-stream throughput, and a DSP-component breakdown on the host machine. Run it on the actual deployment host to see real numbers (the per-call cost varies ~5-10× between laptop CPUs and server-class x86 / ARM):
+
+```
+python scripts/benchmark.py
+python scripts/benchmark.py --json results.json
+python scripts/benchmark.py --multi-streams 1,4,8,16,32 --multi-frames 1000
+python scripts/benchmark.py --skip-dsp    # single + multi-stream only
+```
+
+The script prints a system-info header (Python, NumPy, ORT versions; CPU count), then three sections:
+1. **Single-stream latency** — full per-frame pipeline (`process_frame`), 5 trials, reports the median in µs/frame and the real-time factor (10 ms / per-frame).
+2. **Multi-stream throughput** — N concurrent sessions advancing in lockstep via a thread pool, reports elapsed time and CPU% of one core. Useful for sizing the agent worker (e.g. "with 16 streams we hit 35% of one core on this CPU").
+3. **DSP component breakdown** — analysis, ERB projection, ERB+DF normalization, and synthesis, in isolation from the three ONNX sub-model inferences. Shows how much of the per-frame cost is the numpy frontend vs the ORT calls.
+
+`--json <path>` writes the same numbers in machine-readable form for tracking regressions across machines.
+
+---
+
+## Audio samples
+
+Noisy originals and their Hush-denoised counterparts are in `docs/audio/`. The denoised files are produced by `scripts/process_audio_samples.py` using the same per-frame streaming pipeline the LiveKit frame processor uses in production (one 10 ms frame at a time, continuous GRU state, single session, no resets).
+
+| Original | Denoised |
+|---|---|
+| [`gym.wav`](docs/audio/originals/gym.wav) | [`hush-gym.wav`](docs/audio/hush-gym.wav) |
+| [`krisp-original.wav`](docs/audio/originals/krisp-original.wav) | [`hush-krisp-original.wav`](docs/audio/hush-krisp-original.wav) |
+| [`noproblem_raw.wav`](docs/audio/originals/noproblem_raw.wav) | [`hush-noproblem_raw.wav`](docs/audio/hush-noproblem_raw.wav) |
+| [`taxi-sample.wav`](docs/audio/originals/taxi-sample.wav) | [`hush-taxi-sample.wav`](docs/audio/hush-taxi-sample.wav) |
+
+---
+
+## References
+
+- [pulp-vision/Hush](https://github.com/pulp-vision/Hush) — model architecture and training code
+- [Rikorose/DeepFilterNet](https://github.com/Rikorose/DeepFilterNet) — underlying architecture and DeepFilterLib
+- [Schröter et al., "DeepFilterNet", Interspeech 2023](https://arxiv.org/abs/2305.08227)
+- [LiveKit Agents](https://github.com/livekit/agents)
+
+## License
+
+Apache 2.0

livekit_plugins_hush-0.3.3/README.md

@@ -0,0 +1,169 @@
+# livekit-plugins-hush
+
+[LiveKit](https://livekit.io) noise suppression plugin using the [Hush](https://github.com/pulp-vision/Hush) speech enhancement model. Self-hosted, in-process, no cloud API.
+
+Hush is built on [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet) with an auxiliary separation head for background speaker suppression. Inference uses a pure-numpy DSP frontend for STFT/ERB feature extraction and [ONNX Runtime](https://onnxruntime.ai) for the neural network. No PyTorch, no Rust toolchain, no prebuilt mystery binaries.
+
+---
+
+## Installation
+
+```
+pip install livekit-plugins-hush
+```
+
+Dependencies: `livekit >= 1.0.25`, `livekit-agents >= 1.4.4`, `numpy >= 1.26.0`, `onnxruntime >= 1.17.0`.
+
+---
+
+## Usage
+
+```python
+from livekit.agents import room_io
+from livekit.plugins import hush
+
+await session.start(
+    room_options=room_io.RoomOptions(
+        audio_input=room_io.AudioInputOptions(
+            noise_cancellation=hush.noise_suppression(),
+        ),
+    ),
+)
+```
+
+One instance per session. The ONNX model is loaded once per worker process and shared across instances.
+
+### Parameters
+
+```python
+hush.noise_suppression(
+    strength=0.5,       # wet/dry blend: 0.0 = bypass, 1.0 = full (default 0.5)
+    atten_lim_db=100.0, # max attenuation in dB (100.0 = unlimited)
+    debug_logging=False, # log per-frame RMS every 100 frames
+)
+```
+
+---
+
+## Architecture
+
+The model operates at 16 kHz with 10 ms frames (160 samples, 320-sample FFT). Processing is **per-frame streaming**: one 160-sample frame in, one 160-sample frame out, matching the API shape of the upstream `weya_nc` C library.
+
+The three ONNX sub-models (`enc`, `erb_dec`, `df_dec`) have been re-exported to expose the encoder, ERB decoder, and DF decoder GRU hidden states as I/O. `HushSession` holds those three states (and a 4-frame DF filter history) and threads them through every call, so the GRU has continuous memory across the entire session — same mechanism as the C library.
+
+The pure-numpy STFT runs in streaming mode (`reset=False`), so its analysis/synthesis filter state is carried across frames. The first frame's output is the STFT warmup (effectively zero); algorithmic latency is **1 frame (10 ms)** for the STFT lookahead, on top of GRU-state propagation which is fully causal.
+
+### Signal flow
+
+```
+LiveKit AudioFrame (any rate, any channels)
+  → resample to 16 kHz mono
+  → buffer 10 ms (160 samples) per frame
+  → Pure-numpy streaming STFT (1 frame out, state carried)
+  → ERB features + DF spectral features
+  → ONNX: encoder → ERB decoder + DF decoder (with continuous GRU state)
+  → apply ERB mask + DF complex filter to spectrum
+  → apply atten_lim_db linear blend (matches upstream reference)
+  → Pure-numpy streaming ISTFT
+  → wet/dry blend
+  → upsample, restore channels
+→ AudioFrame
+```
+
+### Model sharing
+
+The ONNX session is a process-level singleton. Each `HushNoiseSuppressor` instance holds only per-stream DF state and audio buffers.
+
+| Per worker | Per session |
+|---|---|
+| ~40 MB (onnxruntime + model) | ~1 MB (DF state + buffers) |
+
+---
+
+## Inference performance
+
+| | Per-frame streaming | Batch (full file) |
+|---|---|---|
+| Frame size | 1 frame (10 ms) | Full audio |
+| Algorithmic latency | 10 ms | N/A (offline) |
+| Per-frame time (single session) | ~0.25 ms | — |
+| Per-frame time (20 concurrent) | ~0.20 ms | — |
+| Real-time factor | 0.025× (per frame) | 0.008× |
+| Throughput | ~4,000 frames/sec per core (single session) | 129× real-time |
+| Concurrent sessions per core | ~390+ at 100 fps | — |
+| Model size | ~9 MB (3 ONNX files) | |
+
+Per-frame time measured on a single ARM64 core with the silero-style
+low-latency ORT config (`intra_op_num_threads=1`, `inter_op_num_threads=1`,
+`ORT_SEQUENTIAL`, no spinning waits). End-to-end `process_frame` takes
+~0.35 ms per 10 ms frame (~28× real-time headroom on a single core).
+Of that, ~0.085 ms is the pure-numpy DSP (STFT/ISTFT + ERB projection
++ per-band EMA) and ~0.27 ms is the three ONNX sub-model inferences
+(encoder + ERB decoder + DF decoder). For typical LiveKit use cases
+(5-20 concurrent calls per agent worker), the plugin is comfortably
+real-time on a single core.
+
+**Concurrency model:** the plugin is GIL-bound. N concurrent sessions
+on one core share the same Python GIL; per-frame work is mostly
+C-level (ORT) which releases the GIL, plus a small amount of pure-numpy
+DSP (STFT + ERB projection + per-band EMA). For higher concurrency
+than ~500 streams per core, run multiple agent worker processes.
+
+**ORT config:** we use the same low-latency session options as the
+upstream [silero VAD plugin](https://github.com/livekit/agents/tree/main/livekit-plugins/livekit-plugins-silero)
+— `intra_op_num_threads=1`, `inter_op_num_threads=1`,
+`execution_mode=ORT_SEQUENTIAL`, no spinning waits. This is ~2× faster
+than ORT defaults for single-stream inference because it avoids the
+per-op thread-pool overhead ORT enables by default.
+
+---
+
+## ONNX model
+
+The ONNX model bundle is from the public PyTorch checkpoint ([weya-ai/hush](https://huggingface.co/weya-ai/hush)). The three sub-models are re-exported with GRU hidden state as I/O using `scripts/export_onnx_stateful.py`; the export script downloads the PyTorch weights from Hugging Face and exports the new ONNX bundle.
+
+Output parity is verified via `scripts/verify_against_pytorch.py`, which compares the per-frame ONNX pipeline output against the original PyTorch model.
+
+## Benchmarking
+
+`scripts/benchmark.py` measures per-frame latency, multi-stream throughput, and a DSP-component breakdown on the host machine. Run it on the actual deployment host to see real numbers (the per-call cost varies ~5-10× between laptop CPUs and server-class x86 / ARM):
+
+```
+python scripts/benchmark.py
+python scripts/benchmark.py --json results.json
+python scripts/benchmark.py --multi-streams 1,4,8,16,32 --multi-frames 1000
+python scripts/benchmark.py --skip-dsp    # single + multi-stream only
+```
+
+The script prints a system-info header (Python, NumPy, ORT versions; CPU count), then three sections:
+1. **Single-stream latency** — full per-frame pipeline (`process_frame`), 5 trials, reports the median in µs/frame and the real-time factor (10 ms / per-frame).
+2. **Multi-stream throughput** — N concurrent sessions advancing in lockstep via a thread pool, reports elapsed time and CPU% of one core. Useful for sizing the agent worker (e.g. "with 16 streams we hit 35% of one core on this CPU").
+3. **DSP component breakdown** — analysis, ERB projection, ERB+DF normalization, and synthesis, in isolation from the three ONNX sub-model inferences. Shows how much of the per-frame cost is the numpy frontend vs the ORT calls.
+
+`--json <path>` writes the same numbers in machine-readable form for tracking regressions across machines.
+
+---
+
+## Audio samples
+
+Noisy originals and their Hush-denoised counterparts are in `docs/audio/`. The denoised files are produced by `scripts/process_audio_samples.py` using the same per-frame streaming pipeline the LiveKit frame processor uses in production (one 10 ms frame at a time, continuous GRU state, single session, no resets).
+
+| Original | Denoised |
+|---|---|
+| [`gym.wav`](docs/audio/originals/gym.wav) | [`hush-gym.wav`](docs/audio/hush-gym.wav) |
+| [`krisp-original.wav`](docs/audio/originals/krisp-original.wav) | [`hush-krisp-original.wav`](docs/audio/hush-krisp-original.wav) |
+| [`noproblem_raw.wav`](docs/audio/originals/noproblem_raw.wav) | [`hush-noproblem_raw.wav`](docs/audio/hush-noproblem_raw.wav) |
+| [`taxi-sample.wav`](docs/audio/originals/taxi-sample.wav) | [`hush-taxi-sample.wav`](docs/audio/hush-taxi-sample.wav) |
+
+---
+
+## References
+
+- [pulp-vision/Hush](https://github.com/pulp-vision/Hush) — model architecture and training code
+- [Rikorose/DeepFilterNet](https://github.com/Rikorose/DeepFilterNet) — underlying architecture and DeepFilterLib
+- [Schröter et al., "DeepFilterNet", Interspeech 2023](https://arxiv.org/abs/2305.08227)
+- [LiveKit Agents](https://github.com/livekit/agents)
+
+## License
+
+Apache 2.0

livekit_plugins_hush-0.3.3/pyproject.toml

@@ -0,0 +1,45 @@
+[build-system]
+requires = ["setuptools>=70", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "livekit-plugins-hush"
+version = "0.3.3"
+description = "Hush noise suppression plugin for LiveKit Agents"
+readme = "README.md"
+license = "Apache-2.0"
+requires-python = ">=3.10"
+keywords = ["livekit", "noise-suppression", "noise-cancellation", "hush", "speech", "audio", "voip"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Multimedia :: Sound/Audio :: Speech",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "livekit>=1.0.25",
+    "livekit-agents>=1.4.4",
+    "numpy>=1.26.0",
+    "onnxruntime>=1.17.0",
+]
+
+[project.urls]
+Homepage = "https://github.com/arpesenti/livekit-plugins-hush"
+Repository = "https://github.com/arpesenti/livekit-plugins-hush"
+"Bug Tracker" = "https://github.com/arpesenti/livekit-plugins-hush/issues"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools.package-data]
+"livekit.plugins.hush" = ["models/*.onnx", "models/*.ini"]
+
+[dependency-groups]
+dev = [
+    "pytest>=9.1.0",
+]

livekit_plugins_hush-0.3.3/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

livekit_plugins_hush-0.3.3/src/livekit/plugins/hush/__init__.py

@@ -0,0 +1,59 @@
+from typing import Optional
+
+from livekit.agents import Plugin
+import logging
+
+from .noise_suppressor import HushNoiseSuppressor
+
+logger = logging.getLogger(__name__)
+
+
+class HushPlugin(Plugin):
+    def __init__(self):
+        super().__init__(
+            title="Hush",
+            version="0.3.3",
+            package="livekit-plugins-hush",
+            logger=logger,
+        )
+
+
+def noise_suppression(
+    model_path: Optional[str] = None,
+    atten_lim_db: float = 100.0,
+    strength: float = 0.5,
+    debug_logging: bool = False,
+) -> HushNoiseSuppressor:
+    """Create a HushNoiseSuppressor instance.
+
+    Pass to ``AudioInputOptions(noise_cancellation=hush.noise_suppression())``.
+
+    Parameters
+    ----------
+    model_path : str, optional
+        Path to directory containing enc.onnx, erb_dec.onnx, df_dec.onnx.
+    atten_lim_db : float
+        Maximum attenuation in dB (default 100.0).
+    strength : float
+        Wet/dry blend factor (default 0.5). 0.0 = bypass, 1.0 = full suppression.
+    debug_logging : bool
+        Log diagnostics every 10 chunks at DEBUG level.
+    """
+    if atten_lim_db < 0:
+        logger.warning(
+            "atten_lim_db=%g is negative; clamping to 0 (no attenuation limit). "
+            "Negative values boost gain instead of limiting attenuation.",
+            atten_lim_db,
+        )
+        atten_lim_db = 0.0
+    return HushNoiseSuppressor(
+        model_path=model_path,
+        atten_lim_db=atten_lim_db,
+        strength=strength,
+        debug_logging=debug_logging,
+    )
+
+
+Plugin.register_plugin(HushPlugin())
+
+__all__ = ["HushNoiseSuppressor", "noise_suppression"]