Preface

A pair of undergraduate students have boldly stepped into the AI realm, unveiling a new AI speech model capable of generating podcast-like audio, rivaling Google's NotebookLM. Despite their limited experience in AI, they've engineered a tool that promises greater control over voice generation, offering diverse script customization options.

Lazy bag

Nari Labs, harnessing Google's TPU technology, released Dia, a model equipped with 1.6 billion parameters. Users can customize speech with disfluencies and clone voices with ease.

Main Body

In the bustling world of AI and synthetic speech technology, two resourceful undergraduates have created waves by introducing an AI model designed to rival Google's NotebookLM. This venture into AI was spearheaded by a Korean duo through Nari Labs, an initiative founded with the vision to enhance speech synthesis technology. The domain of synthetic speech tools is rapidly expanding, currently dominated by substantial industry players such as ElevenLabs. Nevertheless, burgeoning startups are continually challenging the status quo, as demonstrated by this new entrant, underpinned by a significant $398 million in venture capital invested in voice AI technology last year alone.

Toby Kim, who co-founded Nari Labs, reveals that their journey into speech AI commenced merely three months prior, driven by aspirations to construct a model allowing extensive script manipulation and voice control. The duo ingeniously utilized Google's TPU Research Cloud for their project, obtaining access to powerful TPU AI chips at no cost. The resulting model, Dia, comprises a formidable 1.6 billion parameters, enhancing its ability to generate dynamic dialogue from scripts. This model empowers users with the autonomy to personalize speakers' tones, and to incorporate nonverbal elements such as laughter or coughs, enhancing the realism of generated speech.

Parameters, the cornerstone of AI predictive capacity, are pivotal in rendering models like Dia highly effective. More parameters generally signify enhanced performance, a tenet theoretically and practically corroborated by Dia's capabilities. This model is now available through the AI development platforms Hugging Face and GitHub, necessitating a PC with at least 10GB of VRAM for optimal operation. Although it can produce random voices, users can input descriptions to tailor the speech style or employ voice cloning to replicate specific individuals, with impressive fidelity.

Testing conducted by TechCrunch has affirmed Dia's operational efficacy, adeptly sustaining two-way dialogues on varied topics. The qualitative competitiveness of its voice synthesis closely aligns with other contemporary tools, and its user-friendly cloning function has been positively reviewed for simplicity and effectiveness. A sample produced by Dia underscores its potential, although the model remains under scrutiny for its minimalistic safeguards against misuse. The potential for generating misleading or fraudulent content is a significant consideration. Although Nari Labs outlines its ethical usage guidelines, the team disclaims liability for any misuse. Moreover, the training data origins for Dia remain undisclosed, prompting speculation around potential copyright infringements during development, a recurring ethical and legal challenge in AI research.

Notwithstanding these concerns, Kim articulates Nari Labs' ambition to establish a comprehensive synthetic voice platform augmented with social connectivity features, aspiring to enlarge Dia's linguistic repertoire and launch further advanced models. A technical disclosure for Dia is also anticipated, signaling Nari's commitment to transparency and ethical practices in AI technology advancement.

Key Insights Table