Art of Novel Signals: Predicting and Forecasting with High Confidence

The frontier of predictive AI is running into a data wall: the high-quality open text that has powered recent models is largely exhausted, and the two ways around it both have ceilings. Reusing existing data yields little after a few passes, and synthetic data degrades models once it grows past a curated minority of the mix. Neither route creates a genuinely new signal.This SBIR idea seeks to break the wall with signal that was never in the training distribution to begin with: multilingual radio, local news, and community reporting from data-sparse regions. This is a large, almost entirely untapped reservoir of high-value, time-sensitive information that the open web never captured and that synthetic generation cannot manufacture. The central bet of this effort is that conflict and instability signal surfaced from this audio, fed into a temporal-knowledge-graph forecasting model, materially improves geopolitical forecasting precision and warning time in exactly the environments where conventional collection is sparse or denied.The approach has four interlocking parts: a radio data engine, per-language automatic speech recognition (ASR) for predominantly oral languages, a synthetic-data strategy that could make broad language coverage affordable, and the forecasting integration that turns transcribed audio into an early warning. DataPublished scaling work on low-resource ASR (Akera et al., 2025) shows that fine-tuning Whisper Large-v3 (state-of-the-art automatic speech recognition model trained on over 5 million hours of labeled data) reaches usable accuracy at roughly 50 hours of transcribed audio per language and crosses the 10% word-error-rate threshold near 200 hours, with gains flattening beyond that. For this effort, two adjustments are core areas of importance. First, those results that were obtained on clean, single-speaker audio; radio is noisy and multi-speaker, so one should expect higher error rates on raw signal and budget for data curation, not just volume. Second, the 200-hour figure is per language, and the commitment is to be operationally relevant, predominantly oral languages, which are precisely the hardest cases.The collection uses two complementary methods. Where stations stream online, the performer will ingest them directly, which extends the reach far beyond the range of any single receiver. In the low-connectivity environments this program targets, however, many stations never reach the internet at all, and in-region software-defined radio receivers are the only way to capture them. This is precisely why the physical radio listeners matter. The audience is not a fallback, but rather a sole means of reaching signal that no online source can carry. When combined, these two methods make coverage independent of both connectivity and device placement. The performer will further supplement this audio with local news feeds and community audio from social-media channels (Telegram, WhatsApp, etc.) in the target countries. Speech is gated to the roughly 15% of transmissions that contain it, draft transcripts are bootstrapped with Whisper Large-v3, and native-speaker annotators (correcting function). Annotators could be sourced at reasonable costs (around $5/hour) through the performer’s existing in-region network and from diaspora communities; where channels carry human captions, found data further lowers the associated cost.Proposed benchmark A two-dimensional benchmark that pairs word-error-rate with hours of training data, reported per language and tagged by acoustic condition, plus a third axis comparing all-real data against real-plus-synthetic mixes. The headline metric becomes the real transcription hours saved to reach a fixed error rate, which is simultaneously a clean scientific result and a direct cost argument.MetricsAcross the regions of interest (Central and Southeast Asia, East and Northeast Africa, and South America), the current SOA forecasting precision is approximately 80%, and the novel-signal approach is expected to raise this toward 90%. Precision alone is an incomplete measure; a model could score high on precision while still missing many true events, so a strong precision figure could overstate the coverage. As the volume of radio-derived signal grows, the system would gain a more complete picture of the event space, which makes accuracy measurable. The program therefore begins accuracy measurement once sufficient data has been accumulated and improves against that baseline.