A Unity 6 integration for Pocket-TTS.

---

⚠️ Hardware & Platform Support

• Platform: Windows & Android.

---

Installation

Follow these steps exactly to ensure all native dependencies are resolved.

Configure manifest.json

Open your project's Packages/manifest.json and update it to include the scoped registry and the Git dependencies.

{
  "scopedRegistries": [
    {
      "name": "npm",
      "url": "https://registry.npmjs.com",
      "scopes": [
        "com.github.asus4"
      ]
    }
  ],
  "dependencies": {
    "com.github.asus4.onnxruntime": "0.4.2",
    "com.github.asus4.onnxruntime.unity": "0.4.2",
    "ai.lookbe.pockettts": "https://github.com/lookbe/pocket-tts-unity.git",

    ... other dependencies
  }
}

---

Mobile Tuning Guide

Running this model on mobile devices requires balancing Performance (Stuttering) vs. Latency vs. Quality. Use the [PocketTTS Stats] log in the console to profile your device.

The logs will show execution times: [TTS Stats] AR: 50ms | Flow: 20ms | Mimi: 300ms

Also watch for the Real-Time Ratio (RTFx): [TTS RT] Ratio: 0.70x (LAGGING)

Tuning Parameters

You can adjust these settings on the PocketTTSModel component:

1. Diffusion Step (Default: 10)

Controlled by DiffusionStep. This is the most impactful setting for performance.

• Lower (e.g., 2-4):
  • ✅ Drastically reduces computation per frame.
  • ✅ Fixes stuttering on weaker devices.
  • ❌ Lower voice quality: Speech may sound more robotic or have artifacts.
• Higher (e.g., 10+):
  • ✅ High-quality, natural speech.
  • ❌ Heavy CPU usage.

2. Audio Chunk Size (Default: 12)

Controlled by AudioChunkSize. This determines how many frames are generated before decoding audio.

• Larger (e.g., 16-24):
  • ✅ Better throughput: Amortizes the heavy overhead of the Mimi Decoder.
  • ❌ Increased Latency: Users wait longer to hear the first word of a sentence.
• Smaller (e.g., 4-8):
  • ✅ Fast response: Good for conversational agents.
  • ❌ Stuttering Risk: The overhead of frequent decoder calls may cause the pipeline to fall behind real-time.

Example Profiles

Mid-Range Device (e.g., Helio G99):

• Goal: Smooth playback, acceptable latency.
• Settings: DiffusionStep: 3, AudioChunkSize: 16
• Result: No stuttering, ~1.4s startup latency.

High-End Device (e.g., Snapdragon 8 Gen 3):

• Goal: Maximum quality.
• Settings: DiffusionStep: 10, AudioChunkSize: 8
• Result: High-fidelity speech, fast response.

---

Requirements: Models

Choose one language from lookbe/pocket-tts-onnx-v2 and dowload file listed below.

• mimi_encoder.onnx
• mimi_decoder_int8.onnx
• text_conditioner.onnx
• flow_lm_main_int8.onnx
• flow_lm_flow_int8.onnx
• tokenizer.model
• bos_before_voice.npy

---

Credits

• Pocket-TTS Developed by Kyutai Labs – A TTS that fits in your CPU (and pocket).

• onnxruntime-unity Developed by asus4 – ONNX Runtime integration for the Unity engine.