Add Swift API for computing speaker embeddings (#2492)

@@ -9,6 +9,9 @@ ls -lh
 
 
 ./run-test-version.sh
 ./run-test-version.sh
 
 
+./run-compute-speaker-embeddings.sh
+rm -fv *.wav *.onnx
+
 ./run-tts-kitten-en.sh
 ./run-tts-kitten-en.sh
 ls -lh
 ls -lh
 rm -rf kitten-*
 rm -rf kitten-*

@@ -21,3 +21,4 @@ test-version
 zipformer-ctc-asr
 zipformer-ctc-asr
 dolphin-ctc-asr
 dolphin-ctc-asr
 tts-kitten-en
 tts-kitten-en
+compute-speaker-embeddings

-/// swfit-api-examples/SherpaOnnx.swift
+/// swift-api-examples/SherpaOnnx.swift
 /// Copyright (c)  2023  Xiaomi Corporation
 /// Copyright (c)  2023  Xiaomi Corporation
 
 
 import Foundation  // For NSString
 import Foundation  // For NSString
@@ -936,6 +936,41 @@ func sherpaOnnxOfflineTtsConfig(
   )
   )
 }
 }
 
 
+class SherpaOnnxWaveWrapper {
+  let wave: UnsafePointer<SherpaOnnxWave>!
+
+  class func readWave(filename: String) -> SherpaOnnxWaveWrapper {
+    let wave = SherpaOnnxReadWave(toCPointer(filename))
+    return SherpaOnnxWaveWrapper(wave: wave)
+  }
+
+  init(wave: UnsafePointer<SherpaOnnxWave>!) {
+    self.wave = wave
+  }
+
+  deinit {
+    if let wave {
+      SherpaOnnxFreeWave(wave)
+    }
+  }
+
+  var numSamples: Int {
+    return Int(wave.pointee.num_samples)
+  }
+
+  var sampleRate: Int {
+    return Int(wave.pointee.sample_rate)
+  }
+
+  var samples: [Float] {
+    if numSamples == 0 {
+      return []
+    } else {
+      return [Float](UnsafeBufferPointer(start: wave.pointee.samples, count: numSamples))
+    }
+  }
+}
+
 class SherpaOnnxGeneratedAudioWrapper {
 class SherpaOnnxGeneratedAudioWrapper {
   /// A pointer to the underlying counterpart in C
   /// A pointer to the underlying counterpart in C
   let audio: UnsafePointer<SherpaOnnxGeneratedAudio>!
   let audio: UnsafePointer<SherpaOnnxGeneratedAudio>!
@@ -960,14 +995,9 @@ class SherpaOnnxGeneratedAudioWrapper {
 
 
   var samples: [Float] {
   var samples: [Float] {
     if let p = audio.pointee.samples {
     if let p = audio.pointee.samples {
-      var samples: [Float] = []
-      for index in 0..<n {
-        samples.append(p[Int(index)])
-      }
-      return samples
+      return [Float](UnsafeBufferPointer(start: p, count: Int(n)))
     } else {
     } else {
-      let samples: [Float] = []
-      return samples
+      return []
     }
     }
   }
   }
 
 
@@ -1432,6 +1462,72 @@ class SherpaOnnxOfflineSpeakerDiarizationWrapper {
   }
   }
 }
 }
 
 
+class SherpaOnnxOnlineStreamWrapper {
+  /// A pointer to the underlying counterpart in C
+  let impl: OpaquePointer!
+  init(impl: OpaquePointer!) {
+    self.impl = impl
+  }
+
+  deinit {
+    if let impl {
+      SherpaOnnxDestroyOnlineStream(impl)
+    }
+  }
+
+  func acceptWaveform(samples: [Float], sampleRate: Int = 16000) {
+    SherpaOnnxOnlineStreamAcceptWaveform(impl, Int32(sampleRate), samples, Int32(samples.count))
+  }
+
+  func inputFinished() {
+    SherpaOnnxOnlineStreamInputFinished(impl)
+  }
+}
+
+class SherpaOnnxSpeakerEmbeddingExtractorWrapper {
+  /// A pointer to the underlying counterpart in C
+  let impl: OpaquePointer!
+
+  init(
+    config: UnsafePointer<SherpaOnnxSpeakerEmbeddingExtractorConfig>!
+  ) {
+    impl = SherpaOnnxCreateSpeakerEmbeddingExtractor(config)
+  }
+
+  deinit {
+    if let impl {
+      SherpaOnnxDestroySpeakerEmbeddingExtractor(impl)
+    }
+  }
+
+  var dim: Int {
+    return Int(SherpaOnnxSpeakerEmbeddingExtractorDim(impl))
+  }
+
+  func createStream() -> SherpaOnnxOnlineStreamWrapper {
+    let newStream = SherpaOnnxSpeakerEmbeddingExtractorCreateStream(impl)
+    return SherpaOnnxOnlineStreamWrapper(impl: newStream)
+  }
+
+  func isReady(stream: SherpaOnnxOnlineStreamWrapper) -> Bool {
+    return SherpaOnnxSpeakerEmbeddingExtractorIsReady(impl, stream.impl) == 1 ? true : false
+  }
+
+  func compute(stream: SherpaOnnxOnlineStreamWrapper) -> [Float] {
+    if !isReady(stream: stream) {
+      return []
+    }
+
+    let p = SherpaOnnxSpeakerEmbeddingExtractorComputeEmbedding(impl, stream.impl)
+
+    defer {
+      SherpaOnnxSpeakerEmbeddingExtractorDestroyEmbedding(p)
+    }
+
+    return [Float](UnsafeBufferPointer(start: p, count: dim))
+  }
+}
+
 func sherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig(model: String = "")
 func sherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig(model: String = "")
   -> SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig
   -> SherpaOnnxOfflineSpeechDenoiserGtcrnModelConfig
 {
 {

+/// swift-api-examples/compute-speaker-embeddings.swift
+/// Copyright (c)  2025  Xiaomi Corporation
+/*
+Please download test files used in this script from
+
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/speaker-recongition-models
+*/
+func cosineSimilarity(_ a: [Float], _ b: [Float]) -> Float {
+  precondition(a.count == b.count, "Vectors must have the same length")
+
+  // Dot product
+  let dotProduct = zip(a, b).reduce(0) { $0 + $1.0 * $1.1 }
+
+  // Magnitudes
+  let magA = sqrt(a.reduce(0) { $0 + $1 * $1 })
+  let magB = sqrt(b.reduce(0) { $0 + $1 * $1 })
+
+  // Avoid division by zero
+  guard magA > 0 && magB > 0 else { return 0 }
+
+  return dotProduct / (magA * magB)
+}
+
+func computeEmbedding(extractor: SherpaOnnxSpeakerEmbeddingExtractorWrapper, waveFilename: String)
+  -> [Float]
+{
+  let audio = SherpaOnnxWaveWrapper.readWave(filename: waveFilename)
+  let stream = extractor.createStream()
+  stream.acceptWaveform(samples: audio.samples, sampleRate: audio.sampleRate)
+  stream.inputFinished()
+  return extractor.compute(stream: stream)
+}
+
+func run() {
+  let model = "./wespeaker_zh_cnceleb_resnet34.onnx"
+  var config = sherpaOnnxSpeakerEmbeddingExtractorConfig(model: model)
+  let extractor = SherpaOnnxSpeakerEmbeddingExtractorWrapper(config: &config)
+  let embedding1 = computeEmbedding(extractor: extractor, waveFilename: "./fangjun-sr-1.wav")
+  let embedding2 = computeEmbedding(extractor: extractor, waveFilename: "./fangjun-sr-2.wav")
+  let embedding3 = computeEmbedding(extractor: extractor, waveFilename: "./leijun-sr-1.wav")
+
+  let score12 = cosineSimilarity(embedding1, embedding2)
+  let score13 = cosineSimilarity(embedding1, embedding3)
+  let score23 = cosineSimilarity(embedding2, embedding3)
+
+  print("Score between spk1 and spk2: \(score12)")
+  print("Score between spk1 and spk3: \(score13)")
+  print("Score between spk2 and spk3: \(score23)")
+}
+
+@main
+struct App {
+  static func main() {
+    run()
+  }
+}

+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -d ../build-swift-macos ]; then
+  echo "Please run ../build-swift-macos.sh first!"
+  exit 1
+fi
+
+if [ ! -f ./wespeaker_zh_cnceleb_resnet34.onnx ]; then
+  echo "Please download the pre-trained model for testing."
+  echo "You can refer to"
+  echo ""
+  echo "https://github.com/k2-fsa/sherpa-onnx/releases/tag/speaker-recongition-models"
+  echo ""
+  echo "for help"
+
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/wespeaker_zh_cnceleb_resnet34.onnx
+fi
+
+if [ ! -f ./fangjun-sr-1.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/fangjun-sr-1.wav
+fi
+
+if [ ! -f ./fangjun-sr-2.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/fangjun-sr-2.wav
+fi
+
+if [ ! -f ./leijun-sr-1.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/leijun-sr-1.wav
+fi
+
+if [ ! -e ./compute-speaker-embeddings ]; then
+  # Note: We use -lc++ to link against libc++ instead of libstdc++
+  swiftc \
+    -lc++ \
+    -I ../build-swift-macos/install/include \
+    -import-objc-header ./SherpaOnnx-Bridging-Header.h \
+    ./compute-speaker-embeddings.swift  ./SherpaOnnx.swift \
+    -L ../build-swift-macos/install/lib/ \
+    -l sherpa-onnx \
+    -l onnxruntime \
+    -o compute-speaker-embeddings
+
+  strip compute-speaker-embeddings
+else
+  echo "./compute-speaker-embeddings exists - skip building"
+fi
+
+export DYLD_LIBRARY_PATH=$PWD/../build-swift-macos/install/lib:$DYLD_LIBRARY_PATH
+./compute-speaker-embeddings