---
tags:
- speaker-recognition
- coreml
- on-device
---

# Speaker ID v2 for TheStage Apple SDK

This repository contains the encrypted CoreML speaker-embedding bundle for the SDK's `SpeakerIDPipeline`. The source model is [PalabraAI/redimnet2](https://github.com/PalabraAI/redimnet2), checkpoint `b0-vox2-lm.pt` at source revision `ffe8f272bfaf6a41efb7814b28d94377b1613ac9`. The exported checkpoint SHA-256 is `4b5a8c00f813ab0d92fc81ea7c504475d8201bed608fb2c6aaba65be723e9c61`.

The release bundle is on the [`v1.5` revision](https://huggingface.co/TheStageAI/speaker-id-v2/tree/v1.5/subscription/macos_ios). It requires TheStage Apple SDK 1.5.0 or newer and includes a password-protected `.thestage` archive containing bundle-bound QLPA-encrypted CoreML engines. The SDK downloads and decrypts the bundle automatically:

```swift
let speakerID = try await SpeakerIDPipeline(
    engines_path: "TheStageAI/speaker-id-v2")
```

This repository distributes the Apple SDK inference artifact, not the upstream training checkpoint. The previous `TheStageAI/redimnet2` SDK bundle remains available at `v1.4`.