Nari Labs has released Dia-1.6B, an open-source text-to-speech model that's generating significant buzz in the AI community for its ability to create remarkably natural-sounding dialogues. What makes this release particularly noteworthy is that it was developed by a small team of just two engineers over the course of three months, yet delivers quality that rivals offerings from much larger companies.
 |
|---|
| GitHub repository for the Dia open-source text-to-speech model developed by Nari Labs |
Natural Dialogue Generation
Unlike traditional text-to-speech (TTS) models that generate each speaker's lines separately and then stitch them together, Dia generates entire conversations in a single pass. This approach results in more natural-sounding dialogue with proper pacing, overlaps, and emotional continuity. Community members have been particularly impressed by the model's ability to produce non-verbal elements like laughter, coughing, and throat clearing.
This is really impressive; we're getting close to a dream of mine: the ability to generate proper audiobooks from EPUBs. Not just a robotic single voice for everything, but different, consistent voices for each protagonist.
The quality of Dia's output has surprised many users, with several commenting that the examples sound remarkably human. Some have noted that the demo examples have an almost theatrical quality to them, with one user comparing the style to characters from the TV show The Office. This observation led another commenter to discover that one of the demo examples is indeed based on a scene from that show.
Voice and Emotion Control
A standout feature of Dia is its support for audio prompts, allowing users to condition the output on specific voices or emotional tones. By providing a sample audio clip, users can have the model continue generating speech in that same style. This capability opens up possibilities for consistent character voices in audiobooks, podcasts, and other creative applications.
Some users have reported mixed results with the emotion control features, with one mentioning unexpected artifacts like background music appearing when attempting to specify a happy tone. Despite these occasional quirks, the overall capability to maintain consistent voice characteristics throughout a dialogue appears to be working well.
Hardware Requirements and Accessibility
The full version of Dia currently requires around 10GB of VRAM to run, which places it beyond the reach of users with more modest hardware. However, the developers have indicated they plan to release a quantized version in the future that will reduce these requirements, similar to how Suno's Bark model evolved from needing 16GB to running on just 4GB of VRAM.
Community members have already begun adapting the model for different hardware configurations, with one user successfully getting it running on an M2 Pro MacBook Pro. Another confirmed it works on an M4 chip as well. The developers have mentioned that while GPU support is currently required, CPU support will be added soon.
Open Source Contributions and Future Development
As an open-source project released under the Apache License 2.0, Dia has already begun receiving community contributions. Users have submitted pull requests to improve compatibility with different hardware platforms, and some have discussed Docker implementation strategies.
The developers have outlined several areas for future improvement, including Docker support, inference speed optimization, and quantization for memory efficiency. They've also expressed interest in expanding language support beyond English, which multiple community members have requested.
The release of Dia represents another significant step forward in democratizing access to advanced AI speech synthesis technology. By making their 1.6B parameter model openly available, Nari Labs has provided researchers and developers with a powerful tool that can generate convincingly human dialogue without requiring the resources of a major tech company.
Reference: nari-labs/dia