Generate Speech with Dia

Use [S1]/[S2] for speakers, (laughs), (sighs) etc. for non-verbals. Prepend reference transcript for cloning.

0 / 8192
120

Splitting is automatically disabled if text length is less than 2x Chunk Size. Recommended size: ~100-300 (Default: 120). Make sure that you use Predefined Voices or Voice Cloning mode to ensure voices remain consistent for each chunk.

Generation Parameters

Integer (like 1, 42, 901...) for reproducible results, or -1 for random seeds.

Server Configuration

These settings are saved to config.yaml. Restart the server to apply changes to Server, Model, or Paths sections.

    Reset all settings to their defaults from .env file.

Tips & Tricks for Dia

  • Use **Predefined Voices** for consistent, high-quality output based on provided samples.
  • For **Voice Clone**, upload clean reference audio (.wav/.mp3). Crucially, save the exact transcript of the reference audio as the .txt file with the same name as the audio file. [S1] First speaker [S2] Second speaker or [S1] First speaker if the reference audio file has only one speaker.
  • Use **Random / Dialogue** for multi-speaker text ([S1]/[S2]) or single-speaker generation without cloning.
  • Experiment with **CFG Scale** (higher = more adherence) and **Temperature** (higher = more varied).
  • Use **Generation Seed** integer values like 1, 42, 901... for reproducible results.
  • Enable **Split text** for long inputs (> ~200-300 chars). Note: Using Random/Dialogue mode with splitting and a random seed (-1) may result in different voices per chunk. Use Predefined/Clone or a fixed seed for consistency across chunks.
  • Use the /v1/audio/speech endpoint for OpenAI compatibility.
  • Use the custom /tts endpoint for maximum flexibility and configuring all Dia generation parameters, passing reference audio and transcript information etc.