Add Swift API for VoxLock (#4109)

.github/scripts/test-swift.sh

@@ -10,6 +10,11 @@ ls -lh
 
 ./run-test-version.sh
 
+./run-speech-enhancement-voxlock.sh
+./run-speech-enhancement-voxlock-with-voice-sample.sh
+./run-online-speech-enhancement-voxlock.sh
+./run-online-speech-enhancement-voxlock-with-voice-sample.sh
+
 ./run-tts-pocket-en.sh
 ls -lh
 rm -rf sherpa-onnx-pocket-*

.github/workflows/swift.yaml

@@ -84,3 +84,28 @@ jobs:
             swift-api-examples/test-*-lexicon.wav
             swift-api-examples/generated-*.wav
           if-no-files-found: warn
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: speech-enhancement-voxlock-${{ matrix.os }}
+          # each run-*.sh keeps the input and output wavs in wav/<example>
+          path: swift-api-examples/wav/speech-enhancement-voxlock
+          if-no-files-found: warn
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: speech-enhancement-voxlock-with-voice-sample-${{ matrix.os }}
+          path: swift-api-examples/wav/speech-enhancement-voxlock-with-voice-sample
+          if-no-files-found: warn
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: online-speech-enhancement-voxlock-${{ matrix.os }}
+          path: swift-api-examples/wav/online-speech-enhancement-voxlock
+          if-no-files-found: warn
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: online-speech-enhancement-voxlock-with-voice-sample-${{ matrix.os }}
+          path: swift-api-examples/wav/online-speech-enhancement-voxlock-with-voice-sample
+          if-no-files-found: warn

swift-api-examples/SherpaOnnx.swift

@@ -2076,11 +2076,43 @@ public func sherpaOnnxOfflineSpeechDenoiserDpdfNetModelConfig(
     attenuation_limit_db: attenuationLimitDb)
 }
 
+/// Return an instance of SherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig.
+///
+/// VoxLock keeps the voice of one speaker (the caller) and removes the other
+/// voices. The model is from https://github.com/arbazsiddiqui/voxlock
+///
+/// - Parameters:
+///   - enrollModel: Path to the VoxLock enroll model, e.g., enroll.onnx. It
+///     runs once on a voice sample of the speaker to keep.
+///   - chunkModel: Path to the VoxLock chunk model, e.g., chunk.onnx or
+///     chunk_int8.onnx. It runs on each chunk of the audio.
+///   - voiceSample: 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: the offline denoiser uses the loudest 3 seconds in the
+///     first 15 seconds of the input, and the online denoiser passes the audio
+///     through unchanged until 3 seconds of speech have been heard, then uses
+///     that speech.
+public func sherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(
+  enrollModel: String = "",
+  chunkModel: String = "",
+  voiceSample: String = ""
+) -> SherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig {
+  return SherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(
+    enroll_model: toCPointer(enrollModel),
+    chunk_model: toCPointer(chunkModel),
+    voice_sample: toCPointer(voiceSample))
+}
+
 public func sherpaOnnxOfflineSpeechDenoiserModelConfig(
   gtcrn: SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig =
     sherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig(),
   dpdfnet: SherpaOnnxOfflineSpeechDenoiserDpdfNetModelConfig =
     sherpaOnnxOfflineSpeechDenoiserDpdfNetModelConfig(),
+  voxlock: SherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig =
+    sherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(),
   numThreads: Int = 1,
   provider: String = "cpu",
   debug: Int = 0
@@ -2090,7 +2122,8 @@ public func sherpaOnnxOfflineSpeechDenoiserModelConfig(
     num_threads: Int32(numThreads),
     debug: Int32(debug),
     provider: toCPointer(provider),
-    dpdfnet: dpdfnet
+    dpdfnet: dpdfnet,
+    voxlock: voxlock
   )
 }
 

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

@@ -0,0 +1,99 @@
+// Copyright (c)  2026  Xiaomi Corporation
+//
+// This file shows how to use the online speech enhancement API with VoxLock
+// models 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 example 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. See
+// ./online-speech-enhancement-voxlock.swift if you don't have a voice sample.
+//
+// Please refer to
+// https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+// to download files used in this script
+//
+// 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 AVFoundation
+
+extension AudioBuffer {
+  func array() -> [Float] {
+    return Array(UnsafeBufferPointer(self))
+  }
+}
+
+extension AVAudioPCMBuffer {
+  func array() -> [Float] {
+    return self.audioBufferList.pointee.mBuffers.array()
+  }
+}
+
+func run() {
+  var config = sherpaOnnxOnlineSpeechDenoiserConfig(
+    model: sherpaOnnxOfflineSpeechDenoiserModelConfig(
+      voxlock: sherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(
+        enrollModel: "./sherpa-onnx-voxlock-int8/enroll.onnx",
+        chunkModel: "./sherpa-onnx-voxlock-int8/chunk_int8.onnx",
+        voiceSample: "./sherpa-onnx-voxlock-int8/test_wavs/en-voice-sample-f.wav")))
+
+  let sd = SherpaOnnxOnlineSpeechDenoiserWrapper(config: &config)
+
+  let fileURL: NSURL =
+    NSURL(fileURLWithPath: "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav")
+  let audioFile = try! AVAudioFile(forReading: fileURL as URL)
+
+  let audioFormat = audioFile.processingFormat
+  assert(audioFormat.sampleRate == 16000)
+  assert(audioFormat.channelCount == 1)
+  assert(audioFormat.commonFormat == AVAudioCommonFormat.pcmFormatFloat32)
+
+  let audioFrameCount = UInt32(audioFile.length)
+  let audioFileBuffer = AVAudioPCMBuffer(pcmFormat: audioFormat, frameCapacity: audioFrameCount)
+
+  try! audioFile.read(into: audioFileBuffer!)
+  let samples: [Float]! = audioFileBuffer?.array()
+
+  var enhanced: [Float] = []
+  let frameShift = sd.frameShiftInSamples
+
+  let begin = Date()
+  var start = 0
+  while start < samples.count {
+    let end = min(start + frameShift, samples.count)
+    let audio = sd.run(samples: Array(samples[start..<end]), sampleRate: Int(audioFormat.sampleRate))
+    enhanced.append(contentsOf: audio.samples)
+    start = end
+  }
+
+  enhanced.append(contentsOf: sd.flush().samples)
+  let elapsedSeconds = Date().timeIntervalSince(begin)
+
+  let filename = "voxlock-enhanced-online-with-voice-sample-16k.wav"
+  _ = enhanced.withUnsafeBufferPointer { p in
+    SherpaOnnxWriteWave(
+      p.baseAddress,
+      Int32(enhanced.count),
+      Int32(sd.sampleRate),
+      toCPointer(filename))
+  }
+  print("\nSaved to:\(filename)")
+
+  let duration = Double(samples.count) / audioFormat.sampleRate
+  print(String(format: "Elapsed seconds: %.3f", elapsedSeconds))
+  print(String(format: "Audio duration in seconds: %.3f", duration))
+  print(String(format: "RTF: %.3f/%.3f = %.3f", elapsedSeconds, duration, elapsedSeconds / duration))
+}
+
+@main
+struct App {
+  static func main() {
+    run()
+  }
+}

swift-api-examples/online-speech-enhancement-voxlock.swift

@@ -0,0 +1,98 @@
+// Copyright (c)  2026  Xiaomi Corporation
+//
+// This file shows how to use the online speech enhancement API with VoxLock
+// models.
+//
+// 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 example 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. See
+// ./online-speech-enhancement-voxlock-with-voice-sample.swift if you have a
+// voice sample of the speaker.
+//
+// Please refer to
+// https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+// to download files used in this script
+//
+// 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 AVFoundation
+
+extension AudioBuffer {
+  func array() -> [Float] {
+    return Array(UnsafeBufferPointer(self))
+  }
+}
+
+extension AVAudioPCMBuffer {
+  func array() -> [Float] {
+    return self.audioBufferList.pointee.mBuffers.array()
+  }
+}
+
+func run() {
+  var config = sherpaOnnxOnlineSpeechDenoiserConfig(
+    model: sherpaOnnxOfflineSpeechDenoiserModelConfig(
+      voxlock: sherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(
+        enrollModel: "./sherpa-onnx-voxlock-int8/enroll.onnx",
+        chunkModel: "./sherpa-onnx-voxlock-int8/chunk_int8.onnx")))
+
+  let sd = SherpaOnnxOnlineSpeechDenoiserWrapper(config: &config)
+
+  let fileURL: NSURL =
+    NSURL(fileURLWithPath: "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav")
+  let audioFile = try! AVAudioFile(forReading: fileURL as URL)
+
+  let audioFormat = audioFile.processingFormat
+  assert(audioFormat.sampleRate == 16000)
+  assert(audioFormat.channelCount == 1)
+  assert(audioFormat.commonFormat == AVAudioCommonFormat.pcmFormatFloat32)
+
+  let audioFrameCount = UInt32(audioFile.length)
+  let audioFileBuffer = AVAudioPCMBuffer(pcmFormat: audioFormat, frameCapacity: audioFrameCount)
+
+  try! audioFile.read(into: audioFileBuffer!)
+  let samples: [Float]! = audioFileBuffer?.array()
+
+  var enhanced: [Float] = []
+  let frameShift = sd.frameShiftInSamples
+
+  let begin = Date()
+  var start = 0
+  while start < samples.count {
+    let end = min(start + frameShift, samples.count)
+    let audio = sd.run(samples: Array(samples[start..<end]), sampleRate: Int(audioFormat.sampleRate))
+    enhanced.append(contentsOf: audio.samples)
+    start = end
+  }
+
+  enhanced.append(contentsOf: sd.flush().samples)
+  let elapsedSeconds = Date().timeIntervalSince(begin)
+
+  let filename = "voxlock-enhanced-online-16k.wav"
+  _ = enhanced.withUnsafeBufferPointer { p in
+    SherpaOnnxWriteWave(
+      p.baseAddress,
+      Int32(enhanced.count),
+      Int32(sd.sampleRate),
+      toCPointer(filename))
+  }
+  print("\nSaved to:\(filename)")
+
+  let duration = Double(samples.count) / audioFormat.sampleRate
+  print(String(format: "Elapsed seconds: %.3f", elapsedSeconds))
+  print(String(format: "Audio duration in seconds: %.3f", duration))
+  print(String(format: "RTF: %.3f/%.3f = %.3f", elapsedSeconds, duration, elapsedSeconds / duration))
+}
+
+@main
+struct App {
+  static func main() {
+    run()
+  }
+}

swift-api-examples/run-online-speech-enhancement-voxlock-with-voice-sample.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -d ../build-macos ]; then
+  echo "Please run ../build-macos.sh first!"
+  exit 1
+fi
+
+if [ ! -d ./sherpa-onnx-voxlock-int8 ]; then
+  curl -SL -O 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
+  rm sherpa-onnx-voxlock-int8.tar.bz2
+fi
+
+if [ ! -e ./online-speech-enhancement-voxlock-with-voice-sample ]; then
+  # Note: We use -lc++ to link against libc++ instead of libstdc++
+  swiftc \
+    -lc++ \
+    -I ../build-macos/install/include \
+    -import-objc-header ./SherpaOnnx-Bridging-Header.h \
+    ./online-speech-enhancement-voxlock-with-voice-sample.swift ./SherpaOnnx.swift \
+    -L ../build-macos/install/lib/ \
+    -l sherpa-onnx-c-api \
+    -l onnxruntime \
+    -o online-speech-enhancement-voxlock-with-voice-sample
+
+  strip online-speech-enhancement-voxlock-with-voice-sample
+else
+  echo "./online-speech-enhancement-voxlock-with-voice-sample  exists - skip building"
+fi
+
+export DYLD_LIBRARY_PATH=$PWD/../build-macos/install/lib:$DYLD_LIBRARY_PATH
+./online-speech-enhancement-voxlock-with-voice-sample
+
+mkdir -p wav/online-speech-enhancement-voxlock-with-voice-sample
+mv voxlock-enhanced-online-with-voice-sample-16k.wav 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 wav/online-speech-enhancement-voxlock-with-voice-sample/

swift-api-examples/run-online-speech-enhancement-voxlock.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -d ../build-macos ]; then
+  echo "Please run ../build-macos.sh first!"
+  exit 1
+fi
+
+if [ ! -d ./sherpa-onnx-voxlock-int8 ]; then
+  curl -SL -O 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
+  rm sherpa-onnx-voxlock-int8.tar.bz2
+fi
+
+if [ ! -e ./online-speech-enhancement-voxlock ]; then
+  # Note: We use -lc++ to link against libc++ instead of libstdc++
+  swiftc \
+    -lc++ \
+    -I ../build-macos/install/include \
+    -import-objc-header ./SherpaOnnx-Bridging-Header.h \
+    ./online-speech-enhancement-voxlock.swift ./SherpaOnnx.swift \
+    -L ../build-macos/install/lib/ \
+    -l sherpa-onnx-c-api \
+    -l onnxruntime \
+    -o online-speech-enhancement-voxlock
+
+  strip online-speech-enhancement-voxlock
+else
+  echo "./online-speech-enhancement-voxlock  exists - skip building"
+fi
+
+export DYLD_LIBRARY_PATH=$PWD/../build-macos/install/lib:$DYLD_LIBRARY_PATH
+./online-speech-enhancement-voxlock
+
+mkdir -p wav/online-speech-enhancement-voxlock
+mv voxlock-enhanced-online-16k.wav wav/online-speech-enhancement-voxlock/
+cp sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav wav/online-speech-enhancement-voxlock/

swift-api-examples/run-speech-enhancement-voxlock-with-voice-sample.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -d ../build-macos ]; then
+  echo "Please run ../build-macos.sh first!"
+  exit 1
+fi
+
+if [ ! -d ./sherpa-onnx-voxlock-int8 ]; then
+  curl -SL -O 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
+  rm sherpa-onnx-voxlock-int8.tar.bz2
+fi
+
+if [ ! -e ./speech-enhancement-voxlock-with-voice-sample ]; then
+  # Note: We use -lc++ to link against libc++ instead of libstdc++
+  swiftc \
+    -lc++ \
+    -I ../build-macos/install/include \
+    -import-objc-header ./SherpaOnnx-Bridging-Header.h \
+    ./speech-enhancement-voxlock-with-voice-sample.swift ./SherpaOnnx.swift \
+    -L ../build-macos/install/lib/ \
+    -l sherpa-onnx-c-api \
+    -l onnxruntime \
+    -o speech-enhancement-voxlock-with-voice-sample
+
+  strip speech-enhancement-voxlock-with-voice-sample
+else
+  echo "./speech-enhancement-voxlock-with-voice-sample  exists - skip building"
+fi
+
+export DYLD_LIBRARY_PATH=$PWD/../build-macos/install/lib:$DYLD_LIBRARY_PATH
+./speech-enhancement-voxlock-with-voice-sample
+
+mkdir -p wav/speech-enhancement-voxlock-with-voice-sample
+mv voxlock-enhanced-with-voice-sample-16k.wav wav/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 wav/speech-enhancement-voxlock-with-voice-sample/

swift-api-examples/run-speech-enhancement-voxlock.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -d ../build-macos ]; then
+  echo "Please run ../build-macos.sh first!"
+  exit 1
+fi
+
+if [ ! -d ./sherpa-onnx-voxlock-int8 ]; then
+  curl -SL -O 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
+  rm sherpa-onnx-voxlock-int8.tar.bz2
+fi
+
+if [ ! -e ./speech-enhancement-voxlock ]; then
+  # Note: We use -lc++ to link against libc++ instead of libstdc++
+  swiftc \
+    -lc++ \
+    -I ../build-macos/install/include \
+    -import-objc-header ./SherpaOnnx-Bridging-Header.h \
+    ./speech-enhancement-voxlock.swift ./SherpaOnnx.swift \
+    -L ../build-macos/install/lib/ \
+    -l sherpa-onnx-c-api \
+    -l onnxruntime \
+    -o speech-enhancement-voxlock
+
+  strip speech-enhancement-voxlock
+else
+  echo "./speech-enhancement-voxlock  exists - skip building"
+fi
+
+export DYLD_LIBRARY_PATH=$PWD/../build-macos/install/lib:$DYLD_LIBRARY_PATH
+./speech-enhancement-voxlock
+
+mkdir -p wav/speech-enhancement-voxlock
+mv voxlock-enhanced-16k.wav wav/speech-enhancement-voxlock/
+cp sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav wav/speech-enhancement-voxlock/

swift-api-examples/speech-enhancement-voxlock-with-voice-sample.swift

@@ -0,0 +1,86 @@
+// Copyright (c)  2026  Xiaomi Corporation
+//
+// This file shows how to use the speech enhancement API with VoxLock models
+// 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 example 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. See
+// ./speech-enhancement-voxlock.swift if you don't have a voice sample.
+//
+// Please refer to
+// https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+// to download files used in this script
+//
+// 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 AVFoundation
+
+extension AudioBuffer {
+  func array() -> [Float] {
+    return Array(UnsafeBufferPointer(self))
+  }
+}
+
+extension AVAudioPCMBuffer {
+  func array() -> [Float] {
+    return self.audioBufferList.pointee.mBuffers.array()
+  }
+}
+
+func run() {
+  var config = sherpaOnnxOfflineSpeechDenoiserConfig(
+    model: sherpaOnnxOfflineSpeechDenoiserModelConfig(
+      voxlock: sherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(
+        enrollModel: "./sherpa-onnx-voxlock-int8/enroll.onnx",
+        chunkModel: "./sherpa-onnx-voxlock-int8/chunk_int8.onnx",
+        voiceSample: "./sherpa-onnx-voxlock-int8/test_wavs/en-voice-sample-f.wav")))
+
+  let sd = SherpaOnnxOfflineSpeechDenoiserWrapper(config: &config)
+
+  let fileURL: NSURL =
+    NSURL(fileURLWithPath: "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav")
+  let audioFile = try! AVAudioFile(forReading: fileURL as URL)
+
+  let audioFormat = audioFile.processingFormat
+  assert(audioFormat.sampleRate == 16000)
+  assert(audioFormat.channelCount == 1)
+  assert(audioFormat.commonFormat == AVAudioCommonFormat.pcmFormatFloat32)
+
+  let audioFrameCount = UInt32(audioFile.length)
+  let audioFileBuffer = AVAudioPCMBuffer(pcmFormat: audioFormat, frameCapacity: audioFrameCount)
+
+  try! audioFile.read(into: audioFileBuffer!)
+  let array: [Float]! = audioFileBuffer?.array()
+
+  let begin = Date()
+  let audio = sd.run(samples: array, sampleRate: Int(audioFormat.sampleRate))
+  let elapsedSeconds = Date().timeIntervalSince(begin)
+
+  let filename = "voxlock-enhanced-with-voice-sample-16k.wav"
+  let ok = audio.save(filename: filename)
+  if ok == 1 {
+    print("\nSaved to:\(filename)")
+  } else {
+    print("Failed to save to \(filename)")
+  }
+
+  let duration = Double(array.count) / audioFormat.sampleRate
+  print(String(format: "Elapsed seconds: %.3f", elapsedSeconds))
+  print(String(format: "Audio duration in seconds: %.3f", duration))
+  print(String(format: "RTF: %.3f/%.3f = %.3f", elapsedSeconds, duration, elapsedSeconds / duration))
+}
+
+@main
+struct App {
+  static func main() {
+    run()
+  }
+}

swift-api-examples/speech-enhancement-voxlock.swift

@@ -0,0 +1,83 @@
+// Copyright (c)  2026  Xiaomi Corporation
+//
+// This file shows how to use the speech enhancement API with VoxLock models.
+//
+// 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 example 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. See
+// ./speech-enhancement-voxlock-with-voice-sample.swift if you have a voice
+// sample of the speaker.
+//
+// Please refer to
+// https://github.com/k2-fsa/sherpa-onnx/releases/tag/speech-enhancement-models
+// to download files used in this script
+//
+// 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 AVFoundation
+
+extension AudioBuffer {
+  func array() -> [Float] {
+    return Array(UnsafeBufferPointer(self))
+  }
+}
+
+extension AVAudioPCMBuffer {
+  func array() -> [Float] {
+    return self.audioBufferList.pointee.mBuffers.array()
+  }
+}
+
+func run() {
+  var config = sherpaOnnxOfflineSpeechDenoiserConfig(
+    model: sherpaOnnxOfflineSpeechDenoiserModelConfig(
+      voxlock: sherpaOnnxOfflineSpeechDenoiserVoxLockModelConfig(
+        enrollModel: "./sherpa-onnx-voxlock-int8/enroll.onnx",
+        chunkModel: "./sherpa-onnx-voxlock-int8/chunk_int8.onnx")))
+
+  let sd = SherpaOnnxOfflineSpeechDenoiserWrapper(config: &config)
+
+  let fileURL: NSURL =
+    NSURL(fileURLWithPath: "./sherpa-onnx-voxlock-int8/test_wavs/en-combined.wav")
+  let audioFile = try! AVAudioFile(forReading: fileURL as URL)
+
+  let audioFormat = audioFile.processingFormat
+  assert(audioFormat.sampleRate == 16000)
+  assert(audioFormat.channelCount == 1)
+  assert(audioFormat.commonFormat == AVAudioCommonFormat.pcmFormatFloat32)
+
+  let audioFrameCount = UInt32(audioFile.length)
+  let audioFileBuffer = AVAudioPCMBuffer(pcmFormat: audioFormat, frameCapacity: audioFrameCount)
+
+  try! audioFile.read(into: audioFileBuffer!)
+  let array: [Float]! = audioFileBuffer?.array()
+
+  let begin = Date()
+  let audio = sd.run(samples: array, sampleRate: Int(audioFormat.sampleRate))
+  let elapsedSeconds = Date().timeIntervalSince(begin)
+
+  let filename = "voxlock-enhanced-16k.wav"
+  let ok = audio.save(filename: filename)
+  if ok == 1 {
+    print("\nSaved to:\(filename)")
+  } else {
+    print("Failed to save to \(filename)")
+  }
+
+  let duration = Double(array.count) / audioFormat.sampleRate
+  print(String(format: "Elapsed seconds: %.3f", elapsedSeconds))
+  print(String(format: "Audio duration in seconds: %.3f", duration))
+  print(String(format: "RTF: %.3f/%.3f = %.3f", elapsedSeconds, duration, elapsedSeconds / duration))
+}
+
+@main
+struct App {
+  static func main() {
+    run()
+  }
+}