Add Python API for VoxLock (#4102)

.github/scripts/test-python.sh

@@ -14,6 +14,34 @@ source "$SCRIPT_DIR/common.sh"
 log "test version"
 python3 ./python-api-examples/version-test.py
 
+log "test offline speech enhancement (VoxLock)"
+
+download_and_extract https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/sherpa-onnx-voxlock-int8.tar.bz2
+
+python3 ./python-api-examples/offline-speech-enhancement-voxlock.py
+mkdir -p offline-speech-enhancement-voxlock
+mv voxlock-enhanced-16k.wav offline-speech-enhancement-voxlock/
+cp sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav offline-speech-enhancement-voxlock/
+
+python3 ./python-api-examples/offline-speech-enhancement-voxlock-with-voice-sample.py
+mkdir -p offline-speech-enhancement-voxlock-with-voice-sample
+mv voxlock-enhanced-with-voice-sample-16k.wav offline-speech-enhancement-voxlock-with-voice-sample/
+cp sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav sherpa-onnx-voxlock-int8/test_wavs/en-voice-sample-f.wav offline-speech-enhancement-voxlock-with-voice-sample/
+
+python3 ./python-api-examples/online-speech-enhancement-voxlock.py
+mkdir -p online-speech-enhancement-voxlock
+mv voxlock-enhanced-online-16k.wav online-speech-enhancement-voxlock/
+cp sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav online-speech-enhancement-voxlock/
+
+python3 ./python-api-examples/online-speech-enhancement-voxlock-with-voice-sample.py
+mkdir -p online-speech-enhancement-voxlock-with-voice-sample
+mv voxlock-enhanced-online-with-voice-sample-16k.wav online-speech-enhancement-voxlock-with-voice-sample/
+cp sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav sherpa-onnx-voxlock-int8/test_wavs/en-voice-sample-f.wav online-speech-enhancement-voxlock-with-voice-sample/
+
+ls -lh offline-speech-enhancement-voxlock offline-speech-enhancement-voxlock-with-voice-sample online-speech-enhancement-voxlock online-speech-enhancement-voxlock-with-voice-sample
+
+rm -rf sherpa-onnx-voxlock-int8
+
 log "test Cohere Transcribe"
 
 download_and_extract https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-cohere-transcribe-14-lang-int8-2026-04-01.tar.bz2

.github/workflows/run-python-test.yaml

@@ -228,7 +228,7 @@ jobs:
             export PATH="/usr/lib/ccache:/usr/local/opt/ccache/libexec:$PATH"
           fi
           cmake --version
-          export SHERPA_ONNX_CMAKE_ARGS="-G Ninja -DCMAKE_BUILD_TYPE=Release"
+          export SHERPA_ONNX_CMAKE_ARGS="-G Ninja -DCMAKE_BUILD_TYPE=Release -DSHERPA_ONNX_ENABLE_BINARY=OFF"
           export SHERPA_ONNX_MAKE_ARGS="-j 6"
 
           python3 setup.py bdist_wheel
@@ -255,6 +255,7 @@ jobs:
           export CMAKE_CXX_COMPILER_LAUNCHER=ccache
           export PATH="/usr/lib/ccache:/usr/local/opt/ccache/libexec:$PATH"
           cmake --version
+          export SHERPA_ONNX_CMAKE_ARGS="-DSHERPA_ONNX_ENABLE_BINARY=OFF"
 
           python3 -m pip install .
 
@@ -266,6 +267,7 @@ jobs:
             export CMAKE_C_COMPILER_LAUNCHER=sccache
             export CMAKE_CXX_COMPILER_LAUNCHER=sccache
           fi
+          export SHERPA_ONNX_CMAKE_ARGS="-DSHERPA_ONNX_ENABLE_BINARY=OFF"
 
           python3 -m pip install .
 
@@ -291,10 +293,6 @@ jobs:
           unzip ../*.whl
           readelf -d sherpa_onnx/lib/_sherpa_onnx*.so
 
-          echo "----"
-
-          readelf -d sherpa_onnx-*.data/data/bin/sherpa-onnx
-
       - name: Show dependencies - patched (Linux)
         if: runner.os == 'Linux'
         shell: bash
@@ -305,10 +303,6 @@ jobs:
           unzip ../*.whl
           readelf -d sherpa_onnx/lib/_sherpa_onnx*.so
 
-          echo "----"
-
-          readelf -d sherpa_onnx-*.data/data/bin/sherpa-onnx
-
       - name: Test sherpa-onnx (Linux)
         if: runner.os == 'Linux'
         shell: bash
@@ -346,3 +340,29 @@ jobs:
         with:
           name: tts-generated-test-files-${{ matrix.os }}-${{ matrix.python-version }}
           path: tts
+
+      # The VoxLock examples run first in test-python.sh; use if: always() so
+      # their input/output wavs are uploaded even when a later test fails.
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: offline-speech-enhancement-voxlock-${{ matrix.os }}-${{ matrix.python-version }}
+          path: offline-speech-enhancement-voxlock
+
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: offline-speech-enhancement-voxlock-with-voice-sample-${{ matrix.os }}-${{ matrix.python-version }}
+          path: offline-speech-enhancement-voxlock-with-voice-sample
+
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: online-speech-enhancement-voxlock-${{ matrix.os }}-${{ matrix.python-version }}
+          path: online-speech-enhancement-voxlock
+
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: online-speech-enhancement-voxlock-with-voice-sample-${{ matrix.os }}-${{ matrix.python-version }}
+          path: online-speech-enhancement-voxlock-with-voice-sample

python-api-examples/offline-speech-enhancement-voxlock-with-voice-sample.py

@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use the speech enhancement API with VoxLock and a voice
+sample.
+
+VoxLock is a target-speaker voice isolation model. It keeps the voice of one
+speaker, e.g., the caller in a phone call, and removes the other voices.
+
+The model is from https://github.com/arbazsiddiqui/voxlock
+
+This script provides a voice sample of the speaker to keep, so VoxLock isolates
+that speaker from the first sample. The voice sample must have at least 3
+seconds of the speaker's voice; from a longer file, VoxLock uses the loudest
+3 seconds within its first 30 seconds.
+
+If you don't have a voice sample, please see
+./offline-speech-enhancement-voxlock.py instead.
+
+Please download files used in this script from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+
+Example:
+
+ wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/sherpa-onnx-voxlock-int8.tar.bz2
+ tar xvf sherpa-onnx-voxlock-int8.tar.bz2
+"""
+
+import time
+from pathlib import Path
+from typing import Tuple
+
+import numpy as np
+import sherpa_onnx
+import soundfile as sf
+
+
+def create_speech_denoiser():
+    enroll_model = "./sherpa-onnx-voxlock-int8/enroll.onnx"
+    chunk_model = "./sherpa-onnx-voxlock-int8/chunk_int8.onnx"
+    voice_sample = "./sherpa-onnx-voxlock-int8/test_wavs/en-voice-sample-f.wav"
+    if not Path(enroll_model).is_file() or not Path(chunk_model).is_file():
+        raise ValueError(
+            "Please first download the models from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+    if not Path(voice_sample).is_file():
+        raise ValueError(
+            f"{voice_sample} does not exist. You can download it from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    config = sherpa_onnx.OfflineSpeechDenoiserConfig(
+        model=sherpa_onnx.OfflineSpeechDenoiserModelConfig(
+            voxlock=sherpa_onnx.OfflineSpeechDenoiserVoxLockModelConfig(
+                enroll_model=enroll_model,
+                chunk_model=chunk_model,
+                voice_sample=voice_sample,
+            ),
+            debug=False,
+            num_threads=1,
+            provider="cpu",
+        )
+    )
+    if not config.validate():
+        print(config)
+        raise ValueError("Errors in config. Please check previous error logs")
+    return sherpa_onnx.OfflineSpeechDenoiser(config)
+
+
+def load_audio(filename: str) -> Tuple[np.ndarray, int]:
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
+
+
+def main():
+    sd = create_speech_denoiser()
+    test_wave = "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav"
+    if not Path(test_wave).is_file():
+        raise ValueError(
+            f"{test_wave} does not exist. You can download it from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    samples, sample_rate = load_audio(test_wave)
+
+    start = time.time()
+    denoised = sd(samples, sample_rate)
+    end = time.time()
+
+    elapsed_seconds = end - start
+    audio_duration = len(samples) / sample_rate
+    real_time_factor = elapsed_seconds / audio_duration
+
+    sf.write(
+        "./voxlock-enhanced-with-voice-sample-16k.wav",
+        denoised.samples,
+        denoised.sample_rate,
+    )
+    print("Saved to ./voxlock-enhanced-with-voice-sample-16k.wav")
+    print(f"Elapsed seconds: {elapsed_seconds:.3f}")
+    print(f"Audio duration in seconds: {audio_duration:.3f}")
+    print(f"RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}")
+
+
+if __name__ == "__main__":
+    main()

python-api-examples/offline-speech-enhancement-voxlock.py

@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use the speech enhancement API with VoxLock.
+
+VoxLock is a target-speaker voice isolation model. It keeps the voice of one
+speaker, e.g., the caller in a phone call, and removes the other voices.
+
+The model is from https://github.com/arbazsiddiqui/voxlock
+
+This script does not provide a voice sample, so VoxLock takes the loudest
+3 seconds in the first 15 seconds of the input as the speaker to keep.
+
+If you have a voice sample of the speaker, please see
+./offline-speech-enhancement-voxlock-with-voice-sample.py instead.
+
+Please download files used in this script from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+
+Example:
+
+ wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/sherpa-onnx-voxlock-int8.tar.bz2
+ tar xvf sherpa-onnx-voxlock-int8.tar.bz2
+"""
+
+import time
+from pathlib import Path
+from typing import Tuple
+
+import numpy as np
+import sherpa_onnx
+import soundfile as sf
+
+
+def create_speech_denoiser():
+    enroll_model = "./sherpa-onnx-voxlock-int8/enroll.onnx"
+    chunk_model = "./sherpa-onnx-voxlock-int8/chunk_int8.onnx"
+    if not Path(enroll_model).is_file() or not Path(chunk_model).is_file():
+        raise ValueError(
+            "Please first download the models from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    config = sherpa_onnx.OfflineSpeechDenoiserConfig(
+        model=sherpa_onnx.OfflineSpeechDenoiserModelConfig(
+            voxlock=sherpa_onnx.OfflineSpeechDenoiserVoxLockModelConfig(
+                enroll_model=enroll_model,
+                chunk_model=chunk_model,
+                voice_sample="",
+            ),
+            debug=False,
+            num_threads=1,
+            provider="cpu",
+        )
+    )
+    if not config.validate():
+        print(config)
+        raise ValueError("Errors in config. Please check previous error logs")
+    return sherpa_onnx.OfflineSpeechDenoiser(config)
+
+
+def load_audio(filename: str) -> Tuple[np.ndarray, int]:
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
+
+
+def main():
+    sd = create_speech_denoiser()
+    test_wave = "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav"
+    if not Path(test_wave).is_file():
+        raise ValueError(
+            f"{test_wave} does not exist. You can download it from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    samples, sample_rate = load_audio(test_wave)
+
+    start = time.time()
+    denoised = sd(samples, sample_rate)
+    end = time.time()
+
+    elapsed_seconds = end - start
+    audio_duration = len(samples) / sample_rate
+    real_time_factor = elapsed_seconds / audio_duration
+
+    sf.write("./voxlock-enhanced-16k.wav", denoised.samples, denoised.sample_rate)
+    print("Saved to ./voxlock-enhanced-16k.wav")
+    print(f"Elapsed seconds: {elapsed_seconds:.3f}")
+    print(f"Audio duration in seconds: {audio_duration:.3f}")
+    print(f"RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}")
+
+
+if __name__ == "__main__":
+    main()

python-api-examples/online-speech-enhancement-voxlock-with-voice-sample.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use the online speech enhancement API with VoxLock and a
+voice sample.
+
+VoxLock is a target-speaker voice isolation model. It keeps the voice of one
+speaker, e.g., the caller in a phone call, and removes the other voices.
+
+The model is from https://github.com/arbazsiddiqui/voxlock
+
+This script provides a voice sample of the speaker to keep, so VoxLock isolates
+that speaker from the first sample. The voice sample must have at least 3
+seconds of the speaker's voice; from a longer file, VoxLock uses the loudest
+3 seconds within its first 30 seconds.
+
+If you don't have a voice sample, please see
+./online-speech-enhancement-voxlock.py instead.
+
+Please download files used in this script from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+
+Example:
+
+ wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/sherpa-onnx-voxlock-int8.tar.bz2
+ tar xvf sherpa-onnx-voxlock-int8.tar.bz2
+"""
+
+import time
+from pathlib import Path
+
+import numpy as np
+import sherpa_onnx
+import soundfile as sf
+
+
+def create_speech_denoiser():
+    enroll_model = "./sherpa-onnx-voxlock-int8/enroll.onnx"
+    chunk_model = "./sherpa-onnx-voxlock-int8/chunk_int8.onnx"
+    voice_sample = "./sherpa-onnx-voxlock-int8/test_wavs/en-voice-sample-f.wav"
+    if not Path(enroll_model).is_file() or not Path(chunk_model).is_file():
+        raise ValueError(
+            "Please first download the models from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+    if not Path(voice_sample).is_file():
+        raise ValueError(
+            f"{voice_sample} does not exist. You can download it from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    config = sherpa_onnx.OnlineSpeechDenoiserConfig(
+        model=sherpa_onnx.OfflineSpeechDenoiserModelConfig(
+            voxlock=sherpa_onnx.OfflineSpeechDenoiserVoxLockModelConfig(
+                enroll_model=enroll_model,
+                chunk_model=chunk_model,
+                voice_sample=voice_sample,
+            ),
+            debug=False,
+            num_threads=1,
+            provider="cpu",
+        )
+    )
+
+    if not config.validate():
+        print(config)
+        raise ValueError("Errors in config. Please check previous error logs")
+
+    return sherpa_onnx.OnlineSpeechDenoiser(config)
+
+
+def load_audio(filename: str):
+    data, sample_rate = sf.read(filename, always_2d=True, dtype="float32")
+    samples = np.ascontiguousarray(data[:, 0])
+    return samples, sample_rate
+
+
+def main():
+    sd = create_speech_denoiser()
+    test_wave = "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav"
+    if not Path(test_wave).is_file():
+        raise ValueError(
+            f"{test_wave} does not exist. You can download it from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    samples, sample_rate = load_audio(test_wave)
+    frame_shift = sd.frame_shift_in_samples
+    output = []
+
+    start = time.time()
+    for offset in range(0, len(samples), frame_shift):
+        chunk = samples[offset : offset + frame_shift]
+        denoised = sd(chunk, sample_rate)
+        output.append(np.asarray(denoised.samples, dtype=np.float32))
+
+    output.append(np.asarray(sd.flush().samples, dtype=np.float32))
+    end = time.time()
+
+    elapsed_seconds = end - start
+    audio_duration = len(samples) / sample_rate
+    real_time_factor = elapsed_seconds / audio_duration
+
+    enhanced = np.concatenate(output) if output else np.empty(0, dtype=np.float32)
+
+    sf.write(
+        "./voxlock-enhanced-online-with-voice-sample-16k.wav",
+        enhanced,
+        sd.sample_rate,
+    )
+    print("Saved to ./voxlock-enhanced-online-with-voice-sample-16k.wav")
+    print(f"Elapsed seconds: {elapsed_seconds:.3f}")
+    print(f"Audio duration in seconds: {audio_duration:.3f}")
+    print(f"RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}")
+
+
+if __name__ == "__main__":
+    main()

python-api-examples/online-speech-enhancement-voxlock.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use the online speech enhancement API with VoxLock.
+
+VoxLock is a target-speaker voice isolation model. It keeps the voice of one
+speaker, e.g., the caller in a phone call, and removes the other voices.
+
+The model is from https://github.com/arbazsiddiqui/voxlock
+
+This script does not provide a voice sample, so VoxLock passes the audio
+through unchanged until it has heard 3 seconds of speech, takes that speech
+as the speaker to keep, and isolates from then on.
+
+If you have a voice sample of the speaker, please see
+./online-speech-enhancement-voxlock-with-voice-sample.py instead.
+
+Please download files used in this script from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+
+Example:
+
+ wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speech-enhancement-models/sherpa-onnx-voxlock-int8.tar.bz2
+ tar xvf sherpa-onnx-voxlock-int8.tar.bz2
+"""
+
+import time
+from pathlib import Path
+
+import numpy as np
+import sherpa_onnx
+import soundfile as sf
+
+
+def create_speech_denoiser():
+    enroll_model = "./sherpa-onnx-voxlock-int8/enroll.onnx"
+    chunk_model = "./sherpa-onnx-voxlock-int8/chunk_int8.onnx"
+    if not Path(enroll_model).is_file() or not Path(chunk_model).is_file():
+        raise ValueError(
+            "Please first download the models from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    config = sherpa_onnx.OnlineSpeechDenoiserConfig(
+        model=sherpa_onnx.OfflineSpeechDenoiserModelConfig(
+            voxlock=sherpa_onnx.OfflineSpeechDenoiserVoxLockModelConfig(
+                enroll_model=enroll_model,
+                chunk_model=chunk_model,
+                voice_sample="",
+            ),
+            debug=False,
+            num_threads=1,
+            provider="cpu",
+        )
+    )
+
+    if not config.validate():
+        print(config)
+        raise ValueError("Errors in config. Please check previous error logs")
+
+    return sherpa_onnx.OnlineSpeechDenoiser(config)
+
+
+def load_audio(filename: str):
+    data, sample_rate = sf.read(filename, always_2d=True, dtype="float32")
+    samples = np.ascontiguousarray(data[:, 0])
+    return samples, sample_rate
+
+
+def main():
+    sd = create_speech_denoiser()
+    test_wave = "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav"
+    if not Path(test_wave).is_file():
+        raise ValueError(
+            f"{test_wave} does not exist. You can download it from "
+            "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models"
+        )
+
+    samples, sample_rate = load_audio(test_wave)
+    frame_shift = sd.frame_shift_in_samples
+    output = []
+
+    start = time.time()
+    for offset in range(0, len(samples), frame_shift):
+        chunk = samples[offset : offset + frame_shift]
+        denoised = sd(chunk, sample_rate)
+        output.append(np.asarray(denoised.samples, dtype=np.float32))
+
+    output.append(np.asarray(sd.flush().samples, dtype=np.float32))
+    end = time.time()
+
+    elapsed_seconds = end - start
+    audio_duration = len(samples) / sample_rate
+    real_time_factor = elapsed_seconds / audio_duration
+
+    enhanced = np.concatenate(output) if output else np.empty(0, dtype=np.float32)
+
+    sf.write("./voxlock-enhanced-online-16k.wav", enhanced, sd.sample_rate)
+    print("Saved to ./voxlock-enhanced-online-16k.wav")
+    print(f"Elapsed seconds: {elapsed_seconds:.3f}")
+    print(f"Audio duration in seconds: {audio_duration:.3f}")
+    print(f"RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}")
+
+
+if __name__ == "__main__":
+    main()

sherpa-onnx/python/csrc/CMakeLists.txt

@@ -35,6 +35,7 @@ set(srcs
   offline-speech-denoiser-dpdfnet-model-config.cc
   offline-speech-denoiser-gtcrn-model-config.cc
   offline-speech-denoiser-model-config.cc
+  offline-speech-denoiser-voxlock-model-config.cc
   offline-speech-denoiser.cc
   offline-stream.cc
   offline-tdnn-model-config.cc

sherpa-onnx/python/csrc/offline-speech-denoiser-model-config.cc

@@ -9,25 +9,40 @@
 #include "sherpa-onnx/csrc/offline-speech-denoiser-model-config.h"
 #include "sherpa-onnx/python/csrc/offline-speech-denoiser-dpdfnet-model-config.h"
 #include "sherpa-onnx/python/csrc/offline-speech-denoiser-gtcrn-model-config.h"
+#include "sherpa-onnx/python/csrc/offline-speech-denoiser-voxlock-model-config.h"
 
 namespace sherpa_onnx {
 
 void PybindOfflineSpeechDenoiserModelConfig(py::module *m) {
   PybindOfflineSpeechDenoiserDpdfNetModelConfig(m);
   PybindOfflineSpeechDenoiserGtcrnModelConfig(m);
+  PybindOfflineSpeechDenoiserVoxLockModelConfig(m);
 
   using PyClass = OfflineSpeechDenoiserModelConfig;
   py::class_<PyClass>(*m, "OfflineSpeechDenoiserModelConfig")
       .def(py::init<>())
-      .def(py::init<const OfflineSpeechDenoiserGtcrnModelConfig &,
-                    const OfflineSpeechDenoiserDpdfNetModelConfig &, int32_t,
-                    bool, const std::string &>(),
+      .def(py::init([](const OfflineSpeechDenoiserGtcrnModelConfig &gtcrn,
+                       const OfflineSpeechDenoiserDpdfNetModelConfig &dpdfnet,
+                       const OfflineSpeechDenoiserVoxLockModelConfig &voxlock,
+                       int32_t num_threads, bool debug,
+                       const std::string &provider) {
+             PyClass ans;
+             ans.gtcrn = gtcrn;
+             ans.dpdfnet = dpdfnet;
+             ans.voxlock = voxlock;
+             ans.num_threads = num_threads;
+             ans.debug = debug;
+             ans.provider = provider;
+             return ans;
+           }),
            py::arg("gtcrn") = OfflineSpeechDenoiserGtcrnModelConfig{},
            py::arg("dpdfnet") = OfflineSpeechDenoiserDpdfNetModelConfig{},
+           py::arg("voxlock") = OfflineSpeechDenoiserVoxLockModelConfig{},
            py::arg("num_threads") = 1, py::arg("debug") = false,
            py::arg("provider") = "cpu")
       .def_readwrite("gtcrn", &PyClass::gtcrn)
       .def_readwrite("dpdfnet", &PyClass::dpdfnet)
+      .def_readwrite("voxlock", &PyClass::voxlock)
       .def_readwrite("num_threads", &PyClass::num_threads)
       .def_readwrite("debug", &PyClass::debug)
       .def_readwrite("provider", &PyClass::provider)

sherpa-onnx/python/csrc/offline-speech-denoiser-voxlock-model-config.cc

@@ -0,0 +1,72 @@
+// sherpa-onnx/python/csrc/offline-speech-denoiser-voxlock-model-config.cc
+//
+// Copyright (c)  2026  Xiaomi Corporation
+
+#include "sherpa-onnx/python/csrc/offline-speech-denoiser-voxlock-model-config.h"
+
+#include <string>
+
+#include "sherpa-onnx/csrc/offline-speech-denoiser-voxlock-model-config.h"
+
+namespace sherpa_onnx {
+
+static constexpr const char *kVoxLockModelConfigDoc = R"doc(
+Configuration for the VoxLock speech denoiser.
+
+VoxLock keeps the voice of one speaker (the caller) and removes the other
+voices.
+
+The model is from https://github.com/arbazsiddiqui/voxlock
+)doc";
+
+static constexpr const char *kVoxLockEnrollModelDoc = R"doc(
+Path to the VoxLock enroll model, e.g., ``enroll.onnx``.
+
+It runs once on a voice sample of the speaker to keep. Download the models
+from
+https://github.com/arbazsiddiqui/voxlock/releases/tag/sherpa-onnx-models
+)doc";
+
+static constexpr const char *kVoxLockChunkModelDoc = R"doc(
+Path to the VoxLock chunk model, e.g., ``chunk.onnx`` or
+``chunk_int8.onnx``.
+
+It runs on each chunk of the audio.
+)doc";
+
+static constexpr const char *kVoxLockVoiceSampleDoc = R"doc(
+Optional. Path to a wave file with at least 3 seconds of the caller's
+voice. VoxLock keeps this voice and removes the others. From a longer
+file, it uses the loudest 3 seconds within the first 30 seconds. A file
+that is shorter or cannot be read is ignored, as if it were empty.
+
+If empty or ignored, the voice sample is taken from the input audio:
+
+  - Offline denoiser: the loudest 3 seconds in the first 15 seconds of
+    the input are used.
+  - Online denoiser: the audio is passed through unchanged until 3
+    seconds of speech have been heard; that speech is used as the voice
+    sample.
+
+Defaults to ``""``.
+)doc";
+
+void PybindOfflineSpeechDenoiserVoxLockModelConfig(py::module *m) {
+  using PyClass = OfflineSpeechDenoiserVoxLockModelConfig;
+  py::class_<PyClass>(*m, "OfflineSpeechDenoiserVoxLockModelConfig",
+                      kVoxLockModelConfigDoc)
+      .def(py::init<const std::string &, const std::string &,
+                    const std::string &>(),
+           py::arg("enroll_model") = "", py::arg("chunk_model") = "",
+           py::arg("voice_sample") = "")
+      .def_readwrite("enroll_model", &PyClass::enroll_model,
+                     kVoxLockEnrollModelDoc)
+      .def_readwrite("chunk_model", &PyClass::chunk_model,
+                     kVoxLockChunkModelDoc)
+      .def_readwrite("voice_sample", &PyClass::voice_sample,
+                     kVoxLockVoiceSampleDoc)
+      .def("validate", &PyClass::Validate)
+      .def("__str__", &PyClass::ToString);
+}
+
+}  // namespace sherpa_onnx

sherpa-onnx/python/csrc/offline-speech-denoiser-voxlock-model-config.h

@@ -0,0 +1,16 @@
+// sherpa-onnx/python/csrc/offline-speech-denoiser-voxlock-model-config.h
+//
+// Copyright (c)  2026  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SPEECH_DENOISER_VOXLOCK_MODEL_CONFIG_H_
+#define SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SPEECH_DENOISER_VOXLOCK_MODEL_CONFIG_H_
+
+#include "sherpa-onnx/python/csrc/sherpa-onnx.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSpeechDenoiserVoxLockModelConfig(py::module *m);
+
+}
+
+#endif  // SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SPEECH_DENOISER_VOXLOCK_MODEL_CONFIG_H_

sherpa-onnx/python/sherpa_onnx/__init__.py

@@ -47,6 +47,7 @@ from sherpa_onnx.lib._sherpa_onnx import (
     OfflineSpeechDenoiserDpdfNetModelConfig,
     OfflineSpeechDenoiserGtcrnModelConfig,
     OfflineSpeechDenoiserModelConfig,
+    OfflineSpeechDenoiserVoxLockModelConfig,
     OfflineStream,
     OfflineTdnnModelConfig,
     OfflineTransducerModelConfig,