Dia TTS (Nari Labs)

Generate multi-speaker dialog with emotion using Dia TTS by Nari Labs

Dia by Nari Labs is an advanced text-to-speech model that specializes in realistic multi-speaker dialogue. Unlike traditional TTS that handles one speaker at a time, Dia generates natural conversations between multiple speakers with emotion, laughter, hesitation, and other non-verbal cues. At 1.6B parameters, it runs on any 8GB+ GPU.

Key Features

Multi-speaker dialog: Generate conversations between 2+ speakers in one pass
Non-verbal cues: Laughter (laughs), hesitation (sighs), pauses — automatically embedded
Emotional speech: Natural intonation without explicit emotion tags
1.6B parameters: Fits on RTX 3070/3080 (8-10GB VRAM)
Apache 2.0 license: Full commercial use
HuggingFace integration: Works with Transformers library

Requirements

Component

Minimum

Recommended

GPU

RTX 3070 (8GB)

RTX 3080 (10GB)

VRAM

8GB

10GB+

RAM

16GB

32GB

Disk

10GB

15GB

Python

3.9+

3.11

Recommended Clore.ai GPU: RTX 3080 10GB (~$0.2–0.5/day)

Installation

# Option 1: pip install
pip install dia-tts

# Option 2: From source
git clone https://github.com/nari-labs/dia.git
cd dia
pip install -e .

Quick Start

Basic Multi-Speaker Dialog

from dia import Dia

# Load model
model = Dia.from_pretrained("nari-labs/Dia-1.6B")

# Generate multi-speaker conversation
# [S1] = Speaker 1, [S2] = Speaker 2
text = """[S1] Hey, have you tried the new GPU rental platform?
[S2] You mean Clore? Yeah, I rented an RTX 4090 yesterday.
[S1] How was it?
[S2] (laughs) Honestly? Way cheaper than I expected. Like two bucks a day.
[S1] No way. That's... that's actually insane."""

audio = model.generate(text)

# Save to file
import soundfile as sf
sf.write("dialog.wav", audio, samplerate=24000)

With Emotion and Non-Verbal Cues

# Dia automatically handles natural speech patterns
text = """[S1] I just got the results back...
[S2] And? Don't keep me in suspense!
[S1] (sighs) We passed. We actually passed all the tests.
[S2] (laughs) I told you! I told you we'd make it!
[S1] I can't believe it... (laughs) okay, okay, let's celebrate."""

audio = model.generate(text, temperature=0.8)
sf.write("emotional_dialog.wav", audio, samplerate=24000)

Single Speaker

# Works for single speaker too
text = "[S1] Welcome to the Clore AI documentation. In this guide, we'll walk through setting up your first GPU rental and deploying a machine learning model."

audio = model.generate(text)
sf.write("narration.wav", audio, samplerate=24000)

Gradio Web UI

# Launch interactive demo
python -m dia.app --port 7860 --share

# Or manually:
import gradio as gr
from dia import Dia

model = Dia.from_pretrained("nari-labs/Dia-1.6B")

def generate_speech(text):
    audio = model.generate(text)
    return (24000, audio)

demo = gr.Interface(
    fn=generate_speech,
    inputs=gr.Textbox(label="Dialog (use [S1], [S2] tags)", lines=10),
    outputs=gr.Audio(label="Generated Speech"),
    title="Dia TTS — Multi-Speaker Dialog"
)
demo.launch(server_port=7860)

Use Cases

Podcast generation: Create conversational podcasts from scripts
Audiobook dialogs: Generate character conversations with distinct voices
Game dialogue: NPC conversations with natural speech patterns
Training data: Generate diverse speech datasets for ASR training
Chatbot voices: Multi-turn dialog with emotional responses

Tips for Clore.ai Users

RTX 3080 is ideal: 10GB VRAM handles Dia easily at ~$0.2–0.5/day
Batch generation: Process multiple dialogs in a loop to maximize your rental time
Save models to persistent storage: If your Clore instance has persistent disk, cache the model to avoid re-downloading
Temperature 0.7–0.9: Lower = more consistent, higher = more expressive/varied
English only: Dia currently focuses on English — for multilingual, see Qwen3-TTS guide

Troubleshooting

Issue

Solution

CUDA out of memory

Use model.to("cuda", torch_dtype=torch.float16) for half precision

Speakers sound similar

Add more text/context per speaker; try higher temperature

Non-verbal cues ignored

Ensure correct format: (laughs), (sighs) in parentheses

Audio quality low

Increase num_steps parameter if available; ensure 24kHz sample rate