Neva
Neva turns a charades game into a zero-touch dialect speech pipeline: strangers with different mother tongues describe Nano Banana 2 Lite–generated regional picture decks, partners validate meaning in a shared language,
Core trick
Partners confirm each described concept in a shared language, and accepted clips pass quality and de-duplication gates before entering a training corpus.
Description
Neva turns a charades game into a zero-touch dialect speech pipeline: strangers with different mother tongues describe Nano Banana 2 Lite–generated regional picture decks, partners validate meaning in a shared language, and every accepted round is cleaned through automated quality, contamination, and de-duplication gates before it enters an append-only training corpus for fine-tuning—proving that high-throughput creative generation isn’t a prompt box, it’s the engine that keeps play fresh while India-scale language data collects itself.
README excerpt
# Neva **Can the outcome of a communication game supply speech supervision for free?** Neva tests whether *communicative success* — whether a listener recovers the intended concept, and *which* listener recovers it — can replace paid annotation for building dialect-labelled speech corpora. Each record pairs an audio utterance with a deck-owned concept label and a human partner's confirmation that the meaning landed: no transcript, no separate annotation pass. Neva began as a build at the **Google DeepMind Bangalore Hackathon**. The hackathon proved the loop *runs* end-to-end across languages; whether its outcomes are a usable supervision signal — and whether the same loop run *within* a language yields dialect labels for free — is the open research question this repo now tracks. > **Status:** pre-pilot research artifact. This is a paper-and-dataset effort, not a product. > Scope is deliberately bounded. - **Research question & prior work:** [`docs/RESEARCH.md`](docs/RESEARCH.md) - **Phased plan & go/no-go gates:** [`docs/ROADMAP.md`](docs/ROADMAP.md) - **Data availability:** no dataset released yet. Pilot data will be released **CC-BY-4.0** (see [Roadmap](docs/ROADMAP.md), Phase 3). - **Citing Neva:** no preprint yet — see [Citation](#citation). --- ## The research question In a dialect-diverse language (Assamese: Kamrupi, Goalpariya, Central/Nagaon, Eastern/Sivasagar, plus contact varieties Nagamese and Sadri), can gameplay outcomes act as: - **H1 — a quality sign