Modulate launches Velma AI to detect platform risks via acoustic signals

Real-time voice moderation is shifting from simple keyword filtering to native acoustic analysis, helping platforms manage the liability risks of unscripted live audio. As streaming services integrate more social and interactive voice features, the ability to isolate deepfakes and PII in sub-second intervals becomes a core infrastructure requirement rather than a secondary safety feature. This development forces a move away from the traditional 'transcription-then-LLM' workflow, which often loses critical contextual signals like stress or intent. Watch for whether major cloud providers respond by acquiring these specialized audio-native models to bolster their own generic speech-to-text offerings.

Modulate has moved quickly to broaden Velma's reach beyond its gaming origins. In June 2026, the company released its flagship Velma model through a developer API, making it accessible to any developer rather than only enterprise customers, with a showcase planned at Customer Contact Week in Las Vegas. The API expansion positioned Velma as a real-time conversational intelligence layer for fraud prevention, AI agent oversight, contact center operations, and compliance monitoring, all powered by Modulate's Ensemble Listening Model architecture that combines more than 100 specialized models to interpret emotion, tone, intent, and synthetic speech signals.

On the funding side, Modulate raised $25 million in new capital led by Future Ventures in September 2026, with participation from Hyperplane and Lakestar, to accelerate development of its audio-native models and expand its developer and partner ecosystems. The company reported that its technology has analyzed more than 600 million hours of audio across applications spanning fraud and deepfake detection, customer experience, trust and safety, and AI agent supervision. Velma's accuracy claims include 2x greater true-positive detection and 7x fewer false positives compared to traditional LLM-based approaches.

The deepfake detection component represents a specific technical differentiator. Modulate launched Velma Deepfake Detect in March 2026, claiming the #1 position on the Hugging Face Deepfake Speech leaderboard with pricing starting at $0.25 per hour of audio, which the company states is 578x lower cost than the next-best model on that benchmark. The API supports both batch and real-time streaming environments, targeting use cases such as preventing account takeover during customer support calls and flagging synthetic voices during high-value transaction verification. This cost structure is designed to make continuous full-call deepfake monitoring economically viable at enterprise scale, a threshold that previous solutions priced at hundreds of dollars per hour could not cross.