Dia-1.6B TTS : Best Text-to-Dialogue generation AI model
Beats ElevenLabs, Sesame CSM-1B for Conversational AI
Audio AI models are on the rise. We have now seen a number of audio generation models come into the picture this year already. We now have a new model Dia-1.6B, which looks to be better than the ElevenLabs and Sesame CSM 1B models for conversational AI.
https://medium.com/media/891c0824d9ac069656ee089dc0f4499b/href
Data Science in Your Pocket – No Rocket Science
What is Dia-1.6B?
DIA 1.6B is a state-of-the-art 1.6 billion parameter text-to-speech (TTS) model developed by Nari Labs, designed to generate highly realistic and expressive dialogue directly from text transcripts.
Unlike traditional TTS systems that produce rigid or monotonous speech, DIA excels in capturing natural conversational nuances, including dynamic emotion, tone variations, and even nonverbal vocalizations such as laughter, coughing, and throat clearing.
Key Features & Advancements
- Emotion & Tone Control: DIA allows for audio conditioning, meaning users can guide the model’s output by providing reference audio clips to influence the speaker’s emotional delivery and intonation.
- Nonverbal Speech Generation: Beyond standard speech synthesis, DIA can seamlessly interpret tags like (laughs) and reproduce them as natural laughter — a capability absent in models like ElevenLabs and Sesame CSM-1B, which require manual replacements (e.g., “haha”).
- Voice Flexibility: Unlike many TTS models fine-tuned for specific voices, DIA generates varied, randomised voices by default. However, users can fix the seed or provide audio prompts for consistent vocal output.
- Open Sourced: To accelerate innovation in speech synthesis, Nari Labs has released pretrained model checkpoints and inference code, with weights available on Hugging Face.
- Even voice cloning is also supported!
Performance & Comparison
- Early benchmarks suggest that DIA 1.6B outperforms other leading models like Sesame CSM-1B and ElevenLabs in terms of naturalness, expressiveness, and adaptability, particularly in dialogue-heavy scenarios. Though, no official numbers are released.
- While Sesame and ElevenLabs have set high standards in TTS quality, DIA’s larger parameter count, advanced conditioning, and unique ability to handle nonverbal cues give it an edge in producing more context-aware and emotionally nuanced speech.
Current Limitations
As of now, Dia only supports English generation, though future expansions to other languages are anticipated. Researchers and developers can experiment with the model via the provided Hugging Face resources, paving the way for further advancements in AI-driven speech synthesis.
How to use Dia-1.6B for free?
The model is open-sourced and can be accessed quite easily using Google Colab
!pip install git+https://github.com/nari-labs/dia.git
import soundfile as sf
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B")
text = "[S1] Dia is an open weights text to dialogue model (sneezes). [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
output = model.generate(text)
sf.write("simple.mp3", output, 44100)
from IPython.display import Audio
# Upload your MP3 file first or use one from a URL
mp3_path = '/content/simple.mp3' # Change this to your file path
Audio(mp3_path)
Also, if you don’t want to run locally, try here
Dia 1.6B – a Hugging Face Space by nari-labs
Try it, its exceptionally good
Dia-1.6B TTS : Best Text-to-Dialogue generation AI model was originally published in Data Science in Your Pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
