This is an example of using MLX Audio to run local TTS models.

Steps to run:

• Open project in Xcode
• Copy your required files to relevant Resources folder (Kokoro, Orpheus, and Marvis)
• Change project signing in "Signing and Capabilities" project settings
• Run the App

nSpeak framework is embedeed for Kokoro already.

• Required files in Kokoro/Resources folder:
  • kokoro-v1_0.safetensors
  • Voice json files (already in repo)

Implemented and working. Based on Kokoro TTS for iOS. All credit to mlalma for that work!

Uses MLX Swift and eSpeak NG. M1 chip or better is requied.

Files required from MLX Community/Orpheus and MLX Community/Snac-24khz

Currently runs quite slow due to MLX-Swift not letting us compile layers with caching. On an M1 we see a 0.1x processing speed so be patient!

• Required files in Orpheus/Resources folder:
  • orpheus-3b-0.1-ft-4bit.safetensors
  • config.json
  • model.safetensors.index.json
  • snac_model.safetensors
  • snac_config.json
  • tokenizer_config.json
  • tokenizer.json

The full Orpheus functionality is implemented including:

• Voices: tara, leah, jess, leo, dan, mia, zac, zoe
• Expressions: , , , , , , ,

Marvis is an advanced conversational TTS model with streaming support. It uses the Marvis architecture combined with Mimi vocoder for high-quality speech synthesis.

Features:

• Streaming audio generation for real-time TTS
• Two conversational voices: conversational_a and conversational_b
• Downloads model weights automatically on first use from Hugging Face
• Optimized for Apple Silicon with MLX framework

The model runs at 24kHz sample rate and provides natural-sounding conversational speech.