---
license: apache-2.0
language:
- en
- zh
- es
- fr
- hi
- it
- ja
- ko
- pt
tags:
- text-to-speech
- tts
- coreml
- kokoro
- neural-engine
- int8
- palettization
base_model: hexgrad/Kokoro-82M
library_name: coreml
pipeline_tag: text-to-speech
---

# Kokoro-82M CoreML INT8

End-to-end CoreML export of [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M)
with INT8 k-means palettization, optimized for Apple Neural Engine. Requires
iOS 18+ / macOS 15+.

A single `kokoro_5s.mlmodelc` runs the full pipeline (BERT → duration
prediction → fixed-shape alignment → prosody → decoder) in one CoreML
call. G2P (grapheme-to-phoneme) is a separate pair of CoreML models.

## Model

| Parameter | Value |
|-----------|-------|
| Parameters | 82M |
| Precision | INT8 k-means palettization |
| Max audio length | 5 s (200 frames @ 40 fps) |
| Sample rate | 24 kHz |
| Style dimension | 256 |
| Max phonemes per pass | 128 |

## Files

| File | Size | Description |
|------|------|-------------|
| `kokoro_5s.mlmodelc` | 83 MB | Pre-compiled E2E model (pre-compiled, loads directly on-device) |
| `G2PEncoder.mlmodelc` | 0.7 MB | Grapheme-to-phoneme encoder |
| `G2PDecoder.mlmodelc` | 0.8 MB | Grapheme-to-phoneme decoder |
| `voices/` | 0.5 MB | 54 preset voice embeddings (10 languages) |
| `vocab_index.json` | 4 KB | Phoneme vocabulary |
| `g2p_vocab.json` | 4 KB | G2P vocabulary |
| `us_gold.json`, `us_silver.json` | 6 MB | English pronunciation dictionaries |
| `pipeline_config.json` | 4 KB | Swift pipeline config |

## Quality

Compared to the FP16 reference export on a reference 1-second utterance
(`af_heart` voice, 14 phonemes) using the same CoreML inference path:

| Metric | Value |
|--------|-------|
| Predicted duration Δ | 0 frames |
| Output sample count | identical |
| Log-spec distance | **0.42** (close to inaudible) |
| SI-SDR waveform | +0.01 dB |
| Size vs FP16 | **−74%** (83 MB vs 310 MB) |

Because CoreML k-means palettization is not deterministic (scikit-learn's
k-means is unseeded), different exports land at different losses. This
checkpoint was picked from the best of multiple export runs.

## Voices

54 preset voices across 10 languages: English (US/UK), Spanish, French,
Hindi, Italian, Japanese, Korean, Portuguese, Chinese.

## Usage

Add [speech-swift](https://github.com/soniqo/speech-swift) to `Package.swift`:

```swift
.package(url: "https://github.com/soniqo/speech-swift", branch: "main")
```

Then synthesize:

```swift
import KokoroTTS

let tts = try await KokoroTTSModel.fromPretrained(
    modelId: "aufklarer/Kokoro-82M-CoreML-INT8"
)
let audio = try await tts.synthesize(
    "Hello world, this is a Kokoro test.",
    voice: "af_heart"
)
```

CLI:

```bash
swift run audio kokoro "Hello world" --voice af_heart --output out.wav
```

## Source

- Base model: [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) (Apache-2.0)
- Dictionaries and G2P: Apache-2.0

## License

- Model weights: Apache-2.0
- CoreML conversion: Apache-2.0

## Links

- [speech-swift](https://github.com/soniqo/speech-swift) — Apple SDK
- [soniqo.audio](https://soniqo.audio) — website
- [MLX vs CoreML on Apple Silicon — a practical guide](https://blog.ivan.digital/mlx-vs-coreml-on-apple-silicon-a-practical-guide-to-picking-the-right-backend-and-why-you-should-f77ddea7b27a) — related blog post
- [soniqo.audio/blog](https://soniqo.audio/blog) — blog