Nari Dia 1.6B: local setup and dialogue TTS tests

Uploaded on April 22, 2025, this initial test covers Nari Labs' Dia 1.6B text-to-speech model. The setup and capabilities describe that release; the supplied material gives no separate recording date. The presenter describes an Apache 2.0 license, a roughly 6.5 GB model download and no quantized version at the time.

The installation follows the repository's quick start, downloads weights automatically from Hugging Face and opens a Gradio interface. On the presenter's setup, VRAM use reaches about 7.4 GB after loading and roughly 10 GB during generation. He references 3090 hardware and reports rising memory use across later tests. Apple silicon support comes from a repository issue he read, rather than a machine he tested.

The local AI tests start with two-speaker dialogue, then try custom scripts and parenthesized sound cues. One script cuts off before completion; the presenter suggests increasing a generation parameter, but does not confirm a fix. Laughter fares better in his assessment than unusual snorts or angry noises, which produce odd sounds.

Adding a third speaker tag appears to produce another voice, though the presenter explicitly says he does not know whether it is supported. Further experiments read punctuation aloud. These are informal observations, useful for judging dialogue behavior, rather than proof of arbitrary speaker support or a performance benchmark.