aaronslepecky/zant-diarize-fluidaudio

---

language: en license: mit tags: - speaker-diarization - audio - fluidaudio - ios library_name: fluidaudio

ZantOS FluidAudio Diarization Model

Production speaker diarization model for ZantOS iOS app using FluidAudio.

Model Details

• ID: zant-os/zant-diarize-fluidaudio
• Version: 0.6.1
• Backend: FluidAudio + CoreML
• Input: 16kHz mono WAV
• Output: Speaker segments with timestamps and IDs

Performance

Dataset	DER	Speedup	Device
AMI (test)	TBD	50x	iPhone 15 Pro
zant-echo-golden v1.0.0	TBD	50x	iPhone 15 Pro

Usage

This model is consumed via the ZantAMI v0.1 contract.

Swift Integration

import FluidAudio

// Models are downloaded automatically via Swift Package Manager
let engine = ProductionDiarizationEngine.shared
await engine.initialize()

let result = try await engine.diarize(audioURL: url)
// Returns: [SpeakerSegment(start: 0.0, end: 5.3, speaker: "You", confidence: 0.85)]

Model Architecture

FluidAudio uses:
- Segmentation: Voice Activity Detection
- Embedding: WeSpeaker model (256-dimensional)
- Clustering: Hierarchical clustering with cosine similarity

License

MIT License