Text-to-Speech (TTS) Generation

Transformer Lab supports Text-to-Speech (TTS) on MLX (Apple Silicon), CUDA (NVIDIA GPUs), and ROCm (AMD GPUs). This feature lets you convert plain text into natural-sounding speech directly inside Transformer Lab.

How It Works

1. Install the appropriate plugin:

   • Apple Audio MLX Server (for MLX)
   • Unsloth Text-to-Speech Server (for CUDA and ROCm)

2. Select a TTS model in the Foundation tab

3. Switch to the Audio tab

4. Enter text, adjust generation parameters, and generate audio

Supported Model Families

You can start generating audio today with the following models:

MLX (Apple Silicon)

• Kokoro → mlx-community/Kokoro-82M-4bit
• Dia → mlx-community/Dia-1.6B
• Spark → mlx-community/Spark-TTS-0.5B-bf16
• Bark → mlx-community/bark-small
• CSM → mlx-community/csm-1b

CUDA and AMD

• Orpheus → unsloth/orpheus-3b-0.1-ft
• CSM → unsloth/csm-1b

TTS Generation Process

Here's a visual guide to the TTS generation process in Transformer Lab:

This demonstrates the complete workflow from model selection to audio output generation.

Generation Parameters

When generating speech, you’ll see the following parameters:

• Text → The input string to convert into speech
• Sample Rate → Number of audio samples per second (higher = clearer audio)
• Temperature → Controls randomness; lower = consistent, higher = expressive
• Speech Speed → Adjusts pacing of speech (slower = clarity, faster = natural flow)

Some models expose extra controls for more flexibility:

• Audio Cloning → Provide a short reference sample to make the output mimic that voice
• Language → Choose the language for generation (if multilingual support is available)
• Voice → Select a specific voice style or speaker profile offered by the model

Next Steps

Learn how to train your own TTS models:

• Text-to-Speech Training