NeMoSpeaker-iOS

Swift library for speaker embedding extraction using NVIDIA NeMo TitaNet-Small model converted to CoreML.

Features

• Extract 192-dimensional speaker embeddings from audio
• Speaker verification (same/different speaker classification)
• Speaker profiles with embedding aggregation
• Optimized for iOS 16+ and macOS 13+
• Multiple model variants: FP32, FP16, Int8

Requirements

• iOS 16.0+ / macOS 13.0+
• Swift 5.9+

Installation

Swift Package Manager

dependencies: [
    .package(url: "https://github.com/Otosaku/NeMoSpeaker-iOS.git", from: "1.1.0")
]

Download Model

The library does not include the model - you need to download it separately and add to your app bundle.

Download models: Google Drive

Model	Size	Quality	Recommended for
TitaNetSmall.mlmodelc	~27 MB	Best	Development, high accuracy
TitaNetSmall_fp16.mlmodelc	~14 MB	Great	Production (recommended)
TitaNetSmall_int8.mlmodelc	~7 MB	Good	Size-constrained apps

1. Download and unzip the archive
2. Choose the model variant you need
3. Rename to TitaNetSmall.mlmodelc and add to your Xcode project
4. Ensure it's included in "Copy Bundle Resources" build phase

Usage

Extract Speaker Embedding

import NeMoSpeaker

// Get model URL from app bundle
guard let modelURL = Bundle.main.url(forResource: "TitaNetSmall", withExtension: "mlmodelc") else {
    fatalError("Model not found in bundle")
}

// Initialize with model path
let speaker = try NeMoSpeaker(modelURL: modelURL)

// Extract embedding from audio samples (mono, 16kHz, Float32)
let embedding = try speaker.extractEmbedding(samples: audioSamples)

// Embedding is 192-dimensional, L2-normalized
print("Embedding dimension: \(embedding.vector.count)") // 192

Speaker Verification

// Compare two audio samples
let result = try speaker.verify(
    samples1: audioSamples1,
    samples2: audioSamples2,
    threshold: 0.5
)

print("Similarity: \(result.similarity)")      // -1.0 to 1.0
print("Same speaker: \(result.isSameSpeaker)") // true/false

Compare Embeddings Directly

let embedding1 = try speaker.extractEmbedding(samples: samples1)
let embedding2 = try speaker.extractEmbedding(samples: samples2)

// Cosine similarity
let similarity = embedding1.cosineSimilarity(with: embedding2)

// Or use convenience method
let isSame = embedding1.isSameSpeaker(as: embedding2, threshold: 0.5)

Speaker Profiles

// Create a speaker profile
var profile = SpeakerProfile(id: "user_1", embedding: embedding1)

// Add more samples to improve accuracy
profile.addEmbedding(embedding2)
profile.addEmbedding(embedding3)

print("Profile sample count: \(profile.sampleCount)")

// Verify against profile
let result = profile.verify(unknownEmbedding, threshold: 0.5)

Explicit Duration

// Use specific input duration for better control
let embedding = try speaker.extractEmbedding(
    samples: audioSamples,
    duration: .threeSeconds  // 1s, 3s, 5s, or 10s
)

Supported Input Durations

Duration	Audio Samples	Mel Frames
1 sec	16,000	112
3 sec	48,000	304
5 sec	80,000	512
10 sec	160,000	1,008

Audio Requirements

• Sample rate: 16,000 Hz
• Channels: Mono
• Format: Float32

Model Details

• Model: TitaNet-Small (NVIDIA NeMo)
• Embedding dimension: 192
• Variants: FP32 (~27 MB), FP16 (~14 MB), Int8 (~7 MB)
• Original source: NVIDIA NeMo

Threshold Guidelines

Threshold	Use Case
0.4	Lenient (fewer false rejections)
0.5	Balanced (default)
0.6	Strict (fewer false accepts)
0.7+	High security

Example Project

The SpeakerExample folder contains a demo iOS app with:

• Speaker enrollment
• Speaker verification
• Audio comparison
• Live diarization (real-time speaker detection)

To run the example:

1. Open SpeakerExample/SpeakerExample.xcodeproj in Xcode
2. Download model from Google Drive
3. Rename to TitaNetSmall.mlmodelc and drag into the Xcode project
4. Build and run on device or simulator

Dependencies

• NeMoFeatureExtractor-iOS (>= 1.0.5)

License

MIT License