Add Python API for source separation (#2283)

@@ -8,6 +8,32 @@ log() {
   echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
 }
 
+log "test spleeter"
+
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/sherpa-onnx-spleeter-2stems-fp16.tar.bz2
+tar xvf sherpa-onnx-spleeter-2stems-fp16.tar.bz2
+rm sherpa-onnx-spleeter-2stems-fp16.tar.bz2
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/qi-feng-le-zh.wav
+./python-api-examples/offline-source-separation-spleeter.py
+rm -rf sherpa-onnx-spleeter-2stems-fp16
+rm qi-feng-le-zh.wav
+
+log "test UVR"
+
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/UVR_MDXNET_9482.onnx
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/qi-feng-le-zh.wav
+./python-api-examples/offline-source-separation-uvr.py
+rm UVR_MDXNET_9482.onnx
+rm qi-feng-le-zh.wav
+
+mkdir source-separation
+
+mv spleeter-*.wav source-separation
+mv uvr-*.wav source-separation
+
+ls -lh source-separation
+
+
 log "test offline dolphin ctc"
 curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2
 tar xvf sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02.tar.bz2

@@ -99,5 +99,10 @@ jobs:
 
       - uses: actions/upload-artifact@v4
         with:
+          name: source-separation-${{ matrix.os }}-${{ matrix.python-version }}
+          path: ./source-separation
+
+      - uses: actions/upload-artifact@v4
+        with:
           name: tts-generated-test-files-${{ matrix.os }}-${{ matrix.python-version }}
           path: tts

@@ -36,22 +36,18 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          # it fails to install ffmpeg on ubuntu 20.04
-          #
-          # - os: ubuntu-20.04
-          #   python-version: "3.7"
-          # - os: ubuntu-20.04
-          #   python-version: "3.8"
-          # - os: ubuntu-20.04
-          #   python-version: "3.9"
-
-          - os: ubuntu-22.04
+          - os: ubuntu-24.04
+            python-version: "3.8"
+          - os: ubuntu-24.04
+            python-version: "3.9"
+
+          - os: ubuntu-24.04
             python-version: "3.10"
-          - os: ubuntu-22.04
+          - os: ubuntu-24.04
             python-version: "3.11"
-          - os: ubuntu-22.04
+          - os: ubuntu-24.04
             python-version: "3.12"
-          - os: ubuntu-22.04
+          - os: ubuntu-24.04
             python-version: "3.13"
 
     steps:
@@ -81,10 +77,12 @@ jobs:
           python3 -m pip install --upgrade pip numpy pypinyin sentencepiece>=0.1.96 soundfile
           python3 -m pip install wheel twine setuptools
 
-      - name: Install ffmpeg
-        shell: bash
-        run: |
-          sudo apt-get install ffmpeg
+      - uses: afoley587/setup-ffmpeg@main
+        id: setup-ffmpeg
+        with:
+          ffmpeg-version: release
+          architecture: ''
+          github-token: ${{ github.server_url == 'https://github.com' && github.token || '' }}
 
       - name: Install ninja
         shell: bash
@@ -191,5 +189,10 @@ jobs:
 
       - uses: actions/upload-artifact@v4
         with:
+          name: source-separation-${{ matrix.os }}-${{ matrix.python-version }}-whl
+          path: ./source-separation
+
+      - uses: actions/upload-artifact@v4
+        with:
           name: tts-generated-test-files-${{ matrix.os }}-${{ matrix.python-version }}
           path: tts

+#!/usr/bin/env python3
+# Copyright (c)  2025  Xiaomi Corporation
+
+"""
+This file shows how to use spleeter for source separation.
+
+Please first download a spleeter model from
+
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/source-separation-models
+
+The following is an example:
+
+    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/sherpa-onnx-spleeter-2stems-fp16.tar.bz2
+
+Please also download a test file
+
+    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/qi-feng-le-zh.wav
+
+The test wav file is 16-bit encoded with 2 channels. If you have other
+formats, e.g., .mp4 or .mp3, please first use ffmpeg to convert it.
+For instance
+
+    ffmpeg -i your.mp4 -vn -acodec pcm_s16le -ar 44100 -ac 2 out.wav
+
+Then you can use out.wav as input for this example.
+"""
+
+import time
+from pathlib import Path
+
+import numpy as np
+import sherpa_onnx
+import soundfile as sf
+
+
+def create_offline_source_separation():
+    # Please read the help message at the beginning of this file
+    # to download model files
+    vocals = "./sherpa-onnx-spleeter-2stems-fp16/vocals.fp16.onnx"
+    accompaniment = "./sherpa-onnx-spleeter-2stems-fp16/accompaniment.fp16.onnx"
+
+    if not Path(vocals).is_file():
+        raise ValueError(f"{vocals} does not exist.")
+
+    if not Path(accompaniment).is_file():
+        raise ValueError(f"{accompaniment} does not exist.")
+
+    config = sherpa_onnx.OfflineSourceSeparationConfig(
+        model=sherpa_onnx.OfflineSourceSeparationModelConfig(
+            spleeter=sherpa_onnx.OfflineSourceSeparationSpleeterModelConfig(
+                vocals=vocals,
+                accompaniment=accompaniment,
+            ),
+            num_threads=1,
+            debug=False,
+            provider="cpu",
+        )
+    )
+    if not config.validate():
+        raise ValueError("Please check your config.")
+
+    return sherpa_onnx.OfflineSourceSeparation(config)
+
+
+def load_audio():
+    # Please read the help message at the beginning of this file to download
+    # the following wav_file
+    wav_file = "./qi-feng-le-zh.wav"
+    if not Path(wav_file).is_file():
+        raise ValueError(f"{wav_file} does not exist")
+
+    samples, sample_rate = sf.read(wav_file, dtype="float32", always_2d=True)
+    samples = np.transpose(samples)
+    # now samples is of shape (num_channels, num_samples)
+    assert (
+        samples.shape[1] > samples.shape[0]
+    ), f"You should use (num_channels, num_samples). {samples.shape}"
+
+    assert (
+        samples.dtype == np.float32
+    ), f"Expect np.float32 as dtype. Given: {samples.dtype}"
+
+    return samples, sample_rate
+
+
+def main():
+    sp = create_offline_source_separation()
+    samples, sample_rate = load_audio()
+    samples = np.ascontiguousarray(samples)
+
+    start = time.time()
+    output = sp.process(sample_rate=sample_rate, samples=samples)
+    end = time.time()
+
+    print("output.sample_rate", output.sample_rate)
+
+    assert len(output.stems) == 2, len(output.stems)
+
+    vocals = output.stems[0].data
+    non_vocals = output.stems[1].data
+    # vocals.shape (num_channels, num_samples)
+
+    vocals = np.transpose(vocals)
+    non_vocals = np.transpose(non_vocals)
+
+    # vocals.shape (num_samples,num_channels)
+
+    sf.write("./spleeter-vocals.wav", vocals, samplerate=output.sample_rate)
+    sf.write("./spleeter-non-vocals.wav", non_vocals, samplerate=output.sample_rate)
+
+    elapsed_seconds = end - start
+    audio_duration = samples.shape[1] / sample_rate
+    real_time_factor = elapsed_seconds / audio_duration
+
+    print("Saved to ./spleeter-vocals.wav and ./spleeter-non-vocals.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()

+#!/usr/bin/env python3
+# Copyright (c)  2025  Xiaomi Corporation
+
+"""
+This file shows how to use UVR for source separation.
+
+Please first download a UVR model from
+
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/source-separation-models
+
+The following is an example:
+
+    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/UVR_MDXNET_9482.onnx
+
+Please also download a test file
+
+    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/qi-feng-le-zh.wav
+
+The test wav file is 16-bit encoded with 2 channels. If you have other
+formats, e.g., .mp4 or .mp3, please first use ffmpeg to convert it.
+For instance
+
+    ffmpeg -i your.mp4 -vn -acodec pcm_s16le -ar 44100 -ac 2 out.wav
+
+Then you can use out.wav as input for this example.
+"""
+
+import time
+from pathlib import Path
+
+import numpy as np
+import sherpa_onnx
+import soundfile as sf
+
+
+def create_offline_source_separation():
+    # Please read the help message at the beginning of this file
+    # to download model files
+    model = "./UVR_MDXNET_9482.onnx"
+
+    if not Path(model).is_file():
+        raise ValueError(f"{model} does not exist.")
+
+    config = sherpa_onnx.OfflineSourceSeparationConfig(
+        model=sherpa_onnx.OfflineSourceSeparationModelConfig(
+            uvr=sherpa_onnx.OfflineSourceSeparationUvrModelConfig(
+                model=model,
+            ),
+            num_threads=1,
+            debug=False,
+            provider="cpu",
+        )
+    )
+    if not config.validate():
+        raise ValueError("Please check your config.")
+
+    return sherpa_onnx.OfflineSourceSeparation(config)
+
+
+def load_audio():
+    # Please read the help message at the beginning of this file to download
+    # the following wav_file
+    wav_file = "./qi-feng-le-zh.wav"
+    if not Path(wav_file).is_file():
+        raise ValueError(f"{wav_file} does not exist")
+
+    samples, sample_rate = sf.read(wav_file, dtype="float32", always_2d=True)
+    samples = np.transpose(samples)
+    # now samples is of shape (num_channels, num_samples)
+    assert (
+        samples.shape[1] > samples.shape[0]
+    ), f"You should use (num_channels, num_samples). {samples.shape}"
+
+    assert (
+        samples.dtype == np.float32
+    ), f"Expect np.float32 as dtype. Given: {samples.dtype}"
+
+    return samples, sample_rate
+
+
+def main():
+    sp = create_offline_source_separation()
+    samples, sample_rate = load_audio()
+    samples = np.ascontiguousarray(samples)
+
+    print("Started. Please wait")
+    start = time.time()
+    output = sp.process(sample_rate=sample_rate, samples=samples)
+    end = time.time()
+
+    print("output.sample_rate", output.sample_rate)
+
+    assert len(output.stems) == 2, len(output.stems)
+
+    vocals = output.stems[0].data
+    non_vocals = output.stems[1].data
+    # vocals.shape (num_channels, num_samples)
+
+    vocals = np.transpose(vocals)
+    non_vocals = np.transpose(non_vocals)
+
+    # vocals.shape (num_samples,num_channels)
+
+    sf.write("./uvr-vocals.wav", vocals, samplerate=output.sample_rate)
+    sf.write("./uvr-non-vocals.wav", non_vocals, samplerate=output.sample_rate)
+
+    elapsed_seconds = end - start
+    audio_duration = samples.shape[1] / sample_rate
+    real_time_factor = elapsed_seconds / audio_duration
+
+    print("Saved to ./uvr-vocals.wav and ./uvr-non-vocals.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()

@@ -20,6 +20,10 @@ set(srcs
   offline-punctuation.cc
   offline-recognizer.cc
   offline-sense-voice-model-config.cc
+  offline-source-separation-model-config.cc
+  offline-source-separation-spleeter-model-config.cc
+  offline-source-separation-uvr-model-config.cc
+  offline-source-separation.cc
   offline-speech-denoiser-gtcrn-model-config.cc
   offline-speech-denoiser-model-config.cc
   offline-speech-denoiser.cc

@@ -9,6 +9,8 @@
 
 #include "sherpa-onnx/csrc/fast-clustering.h"
 
+#define C_CONTIGUOUS py::detail::npy_api::constants::NPY_ARRAY_C_CONTIGUOUS_
+
 namespace sherpa_onnx {
 
 static void PybindFastClusteringConfig(py::module *m) {
@@ -32,6 +34,12 @@ void PybindFastClustering(py::module *m) {
           "__call__",
           [](const PyClass &self,
              py::array_t<float> features) -> std::vector<int32_t> {
+            if (!(C_CONTIGUOUS == (features.flags() & C_CONTIGUOUS))) {
+              throw py::value_error(
+                  "input features should be contiguous. Please use "
+                  "np.ascontiguousarray(features)");
+            }
+
             int num_dim = features.ndim();
             if (num_dim != 2) {
               std::ostringstream os;

@@ -59,14 +59,14 @@ void PybindOfflineRecognizer(py::module *m) {
             return self.CreateStream(hotwords);
           },
           py::arg("hotwords"), py::call_guard<py::gil_scoped_release>())
-      .def("decode_stream", &PyClass::DecodeStream,
+      .def("decode_stream", &PyClass::DecodeStream, py::arg("s"),
            py::call_guard<py::gil_scoped_release>())
       .def(
           "decode_streams",
           [](const PyClass &self, std::vector<OfflineStream *> ss) {
             self.DecodeStreams(ss.data(), ss.size());
           },
-          py::call_guard<py::gil_scoped_release>());
+          py::arg("ss"), py::call_guard<py::gil_scoped_release>());
 }
 
 }  // namespace sherpa_onnx

+// sherpa-onnx/python/csrc/offline-source-separation-model-config.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/python/csrc/offline-source-separation-model-config.h"
+
+#include <string>
+
+#include "sherpa-onnx/csrc/offline-source-separation-model-config.h"
+#include "sherpa-onnx/python/csrc/offline-source-separation-spleeter-model-config.h"
+#include "sherpa-onnx/python/csrc/offline-source-separation-uvr-model-config.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparationModelConfig(py::module *m) {
+  PybindOfflineSourceSeparationSpleeterModelConfig(m);
+  PybindOfflineSourceSeparationUvrModelConfig(m);
+
+  using PyClass = OfflineSourceSeparationModelConfig;
+  py::class_<PyClass>(*m, "OfflineSourceSeparationModelConfig")
+      .def(py::init<const OfflineSourceSeparationSpleeterModelConfig &,
+                    const OfflineSourceSeparationUvrModelConfig &, int32_t,
+                    bool, const std::string &>(),
+           py::arg("spleeter") = OfflineSourceSeparationSpleeterModelConfig{},
+           py::arg("uvr") = OfflineSourceSeparationUvrModelConfig{},
+           py::arg("num_threads") = 1, py::arg("debug") = false,
+           py::arg("provider") = "cpu")
+      .def_readwrite("spleeter", &PyClass::spleeter)
+      .def_readwrite("uvr", &PyClass::uvr)
+      .def_readwrite("num_threads", &PyClass::num_threads)
+      .def_readwrite("debug", &PyClass::debug)
+      .def_readwrite("provider", &PyClass::provider)
+      .def("validate", &PyClass::Validate)
+      .def("__str__", &PyClass::ToString);
+}
+
+}  // namespace sherpa_onnx

+// sherpa-onnx/python/csrc/offline-source-separation-model-config.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_MODEL_CONFIG_H_
+#define SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_MODEL_CONFIG_H_
+
+#include "sherpa-onnx/python/csrc/sherpa-onnx.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparationModelConfig(py::module *m);
+
+}
+
+#endif  // SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_MODEL_CONFIG_H_

+// sherpa-onnx/python/csrc/offline-source-separation-spleeter-model-config.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/python/csrc/offline-source-separation-spleeter-model-config.h"
+
+#include <string>
+
+#include "sherpa-onnx/csrc/offline-source-separation-spleeter-model-config.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparationSpleeterModelConfig(py::module *m) {
+  using PyClass = OfflineSourceSeparationSpleeterModelConfig;
+  py::class_<PyClass>(*m, "OfflineSourceSeparationSpleeterModelConfig")
+      .def(py::init<const std::string &, const std::string &>(),
+           py::arg("vocals") = "", py::arg("accompaniment") = "")
+      .def_readwrite("vocals", &PyClass::vocals)
+      .def_readwrite("accompaniment", &PyClass::accompaniment)
+      .def("validate", &PyClass::Validate)
+      .def("__str__", &PyClass::ToString);
+}
+
+}  // namespace sherpa_onnx

+// sherpa-onnx/python/csrc/offline-source-separation-spleeter-model-config.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_SPLEETER_MODEL_CONFIG_H_
+#define SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_SPLEETER_MODEL_CONFIG_H_
+
+#include "sherpa-onnx/python/csrc/sherpa-onnx.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparationSpleeterModelConfig(py::module *m);
+
+}
+
+#endif  // SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_SPLEETER_MODEL_CONFIG_H_

+// sherpa-onnx/python/csrc/offline-source-separation-uvr-model-config.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/python/csrc/offline-source-separation-uvr-model-config.h"
+
+#include <string>
+
+#include "sherpa-onnx/csrc/offline-source-separation-uvr-model-config.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparationUvrModelConfig(py::module *m) {
+  using PyClass = OfflineSourceSeparationUvrModelConfig;
+  py::class_<PyClass>(*m, "OfflineSourceSeparationUvrModelConfig")
+      .def(py::init<const std::string &>(), py::arg("model") = "")
+      .def_readwrite("model", &PyClass::model)
+      .def("validate", &PyClass::Validate)
+      .def("__str__", &PyClass::ToString);
+}
+
+}  // namespace sherpa_onnx

+// sherpa-onnx/python/csrc/offline-source-separation-uvr-model-config.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_UVR_MODEL_CONFIG_H_
+#define SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_UVR_MODEL_CONFIG_H_
+
+#include "sherpa-onnx/python/csrc/sherpa-onnx.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparationUvrModelConfig(py::module *m);
+
+}
+
+#endif  // SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_UVR_MODEL_CONFIG_H_

+// sherpa-onnx/python/csrc/offline-source-separation-config.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/offline-source-separation.h"
+
+#include <string>
+
+#include "sherpa-onnx/python/csrc/offline-source-separation-model-config.h"
+#include "sherpa-onnx/python/csrc/offline-source-separation.h"
+
+#define C_CONTIGUOUS py::detail::npy_api::constants::NPY_ARRAY_C_CONTIGUOUS_
+
+namespace sherpa_onnx {
+
+static void PybindOfflineSourceSeparationConfig(py::module *m) {
+  PybindOfflineSourceSeparationModelConfig(m);
+
+  using PyClass = OfflineSourceSeparationConfig;
+  py::class_<PyClass>(*m, "OfflineSourceSeparationConfig")
+      .def(py::init<const OfflineSourceSeparationModelConfig &>(),
+           py::arg("model") = OfflineSourceSeparationModelConfig{})
+      .def_readwrite("model", &PyClass::model)
+      .def("validate", &PyClass::Validate)
+      .def("__str__", &PyClass::ToString);
+}
+
+static void PybindMultiChannelSamples(py::module *m) {
+  using PyClass = MultiChannelSamples;
+
+  py::class_<PyClass>(*m, "MultiChannelSamples")
+      .def_property_readonly("data", [](PyClass &self) -> py::object {
+        // if data is not empty, return a float array of
+        // shape (num_channels, num_samples)
+        int32_t num_channels = self.data.size();
+        if (num_channels == 0) {
+          return py::none();
+        }
+
+        int32_t num_samples = self.data[0].size();
+        if (num_samples == 0) {
+          return py::none();
+        }
+
+        py::array_t<float> ans({num_channels, num_samples});
+
+        py::buffer_info buf = ans.request();
+        auto p = static_cast<float *>(buf.ptr);
+
+        for (int32_t i = 0; i != num_channels; ++i) {
+          std::copy(self.data[i].begin(), self.data[i].end(),
+                    p + i * num_samples);
+        }
+
+        return ans;
+      });
+}
+
+static void PybindOfflineSourceSeparationOutput(py::module *m) {
+  using PyClass = OfflineSourceSeparationOutput;
+  py::class_<PyClass>(*m, "OfflineSourceSeparationOutput")
+      .def_property_readonly(
+          "sample_rate", [](const PyClass &self) { return self.sample_rate; })
+      .def_property_readonly("stems",
+                             [](const PyClass &self) { return self.stems; });
+}
+
+void PybindOfflineSourceSeparation(py::module *m) {
+  PybindOfflineSourceSeparationConfig(m);
+  PybindOfflineSourceSeparationOutput(m);
+
+  PybindMultiChannelSamples(m);
+
+  using PyClass = OfflineSourceSeparation;
+  py::class_<PyClass>(*m, "OfflineSourceSeparation")
+      .def(py::init<const OfflineSourceSeparationConfig &>(),
+           py::arg("config") = OfflineSourceSeparationConfig{})
+      .def(
+          "process",
+          [](const PyClass &self, int32_t sample_rate,
+             const py::array_t<float> &samples) {
+            if (!(C_CONTIGUOUS == (samples.flags() & C_CONTIGUOUS))) {
+              throw py::value_error(
+                  "input samples should be contiguous. Please use "
+                  "np.ascontiguousarray(samples)");
+            }
+
+            int num_dim = samples.ndim();
+            if (samples.ndim() != 2) {
+              std::ostringstream os;
+              os << "Expect an array of 2 dimensions [num_channels x "
+                    "num_samples]. "
+                    "Given dim: "
+                 << num_dim << "\n";
+              throw py::value_error(os.str());
+            }
+
+            // if num_samples is less than 10, it is very likely the user
+            // has swapped num_channels and num_samples.
+            if (samples.shape(1) < 10) {
+              std::ostringstream os;
+              os << "Expect an array of 2 dimensions [num_channels x "
+                    "num_samples]. "
+                    "Given ["
+                 << samples.shape(0) << " x " << samples.shape(1) << "]"
+                 << "\n";
+              throw py::value_error(os.str());
+            }
+
+            int32_t num_channels = samples.shape(0);
+            int32_t num_samples = samples.shape(1);
+            const float *p = samples.data();
+
+            OfflineSourceSeparationInput input;
+
+            input.samples.data.resize(num_channels);
+            input.sample_rate = sample_rate;
+
+            for (int32_t i = 0; i != num_channels; ++i) {
+              input.samples.data[i] = {p + i * num_samples,
+                                       p + (i + 1) * num_samples};
+            }
+
+            pybind11::gil_scoped_release release;
+
+            return self.Process(input);
+          },
+          py::arg("sample_rate"), py::arg("samples"),
+          "samples is of shape (num_channels, num-samples) with dtype "
+          "np.float32");
+}
+
+}  // namespace sherpa_onnx

+// sherpa-onnx/python/csrc/offline-source-separation-config.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_CONFIG_H_
+#define SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_CONFIG_H_
+
+#include "sherpa-onnx/python/csrc/sherpa-onnx.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineSourceSeparation(py::module *m);
+
+}
+
+#endif  // SHERPA_ONNX_PYTHON_CSRC_OFFLINE_SOURCE_SEPARATION_CONFIG_H_

@@ -47,6 +47,7 @@ void PybindOfflineSpeechDenoiser(py::module *m) {
              int32_t sample_rate) {
             return self.Run(samples.data(), samples.size(), sample_rate);
           },
+          py::arg("samples"), py::arg("sample_rate"),
           py::call_guard<py::gil_scoped_release>())
       .def(
           "run",
@@ -54,6 +55,7 @@ void PybindOfflineSpeechDenoiser(py::module *m) {
              int32_t sample_rate) {
             return self.Run(samples.data(), samples.size(), sample_rate);
           },
+          py::arg("samples"), py::arg("sample_rate"),
           py::call_guard<py::gil_scoped_release>())
       .def_property_readonly("sample_rate", &PyClass::GetSampleRate);
 }

@@ -109,19 +109,20 @@ void PybindOnlineRecognizer(py::module *m) {
           py::arg("hotwords"), py::call_guard<py::gil_scoped_release>())
       .def("is_ready", &PyClass::IsReady,
            py::call_guard<py::gil_scoped_release>())
-      .def("decode_stream", &PyClass::DecodeStream,
+      .def("decode_stream", &PyClass::DecodeStream, py::arg("s"),
            py::call_guard<py::gil_scoped_release>())
       .def(
           "decode_streams",
           [](PyClass &self, std::vector<OnlineStream *> ss) {
             self.DecodeStreams(ss.data(), ss.size());
           },
-          py::call_guard<py::gil_scoped_release>())
-      .def("get_result", &PyClass::GetResult,
+          py::arg("ss"), py::call_guard<py::gil_scoped_release>())
+      .def("get_result", &PyClass::GetResult, py::arg("s"),
            py::call_guard<py::gil_scoped_release>())
-      .def("is_endpoint", &PyClass::IsEndpoint,
+      .def("is_endpoint", &PyClass::IsEndpoint, py::arg("s"),
            py::call_guard<py::gil_scoped_release>())
-      .def("reset", &PyClass::Reset, py::call_guard<py::gil_scoped_release>());
+      .def("reset", &PyClass::Reset, py::arg("s"),
+           py::call_guard<py::gil_scoped_release>());
 }
 
 }  // namespace sherpa_onnx

@@ -17,6 +17,7 @@
 #include "sherpa-onnx/python/csrc/offline-model-config.h"
 #include "sherpa-onnx/python/csrc/offline-punctuation.h"
 #include "sherpa-onnx/python/csrc/offline-recognizer.h"
+#include "sherpa-onnx/python/csrc/offline-source-separation.h"
 #include "sherpa-onnx/python/csrc/offline-speech-denoiser.h"
 #include "sherpa-onnx/python/csrc/offline-stream.h"
 #include "sherpa-onnx/python/csrc/online-ctc-fst-decoder-config.h"
@@ -110,6 +111,7 @@ PYBIND11_MODULE(_sherpa_onnx, m) {
 
   PybindAlsa(&m);
   PybindOfflineSpeechDenoiser(&m);
+  PybindOfflineSourceSeparation(&m);
 }
 
 }  // namespace sherpa_onnx

@@ -11,6 +11,11 @@ from _sherpa_onnx import (
     OfflinePunctuation,
     OfflinePunctuationConfig,
     OfflinePunctuationModelConfig,
+    OfflineSourceSeparation,
+    OfflineSourceSeparationConfig,
+    OfflineSourceSeparationModelConfig,
+    OfflineSourceSeparationSpleeterModelConfig,
+    OfflineSourceSeparationUvrModelConfig,
     OfflineSpeakerDiarization,
     OfflineSpeakerDiarizationConfig,
     OfflineSpeakerDiarizationResult,