Dia — Realistic Dialogue Text-to-Speech Model

What it is

Dia is a 1.6B parameter text-to-speech model built by Nari Labs. It generates realistic dialogue audio from text transcripts, supporting multiple speakers in a single generation, non-verbal sounds like laughter and sighs, and voice cloning from reference audio.

Dia targets podcast producers, content creators, and developers building conversational AI interfaces who need natural-sounding multi-speaker audio without recording studios.

How it saves time or tokens

Traditional multi-speaker TTS requires generating each speaker separately and splicing audio. Dia produces a complete multi-speaker conversation in a single pass. The transcript format uses speaker tags like [S1] and [S2], so you write one script and get one audio file with distinct voices.

Voice cloning with a short reference clip eliminates the need for voice actor sessions for consistent character voices.

How to use

1. Install Dia: pip install dia-tts
2. Prepare a transcript with speaker tags and optional non-verbal annotations
3. Run generation with the CLI or Python API
4. For voice cloning, provide a reference audio file for each speaker

Example

from dia import Dia

model = Dia('nari-labs/dia-1.6b')

transcript = '''
[S1] Have you tried the new model?
[S2] (laughs) Yeah, it is surprisingly good.
[S1] Right? The voice quality is way better than I expected.
[S2] I might use it for my podcast intros.
'''

audio = model.generate(
    transcript,
    output_path='dialogue.wav',
    sample_rate=44100
)

The output is a single WAV file with two distinct speaker voices and the laugh rendered naturally.

Related on TokRepo

• Voice tools -- Text-to-speech and voice AI tools
• Content tools -- Content creation and production tools

Common pitfalls

• Voice cloning quality degrades with noisy or short reference clips; use clean audio of at least 10 seconds
• Non-verbal sound tags must match the model's vocabulary; unsupported tags are silently ignored
• Running the 1.6B model requires a GPU with at least 6 GB VRAM; CPU inference is possible but very slow