Modulate Raises $25 Million to Advance Deepfake Detection

* Modulate raised $25 million in funding from Future Ventures, with participation from Hyperplane and Lakestar, totaling $60 million.
* Modulate builds audio-native AI models to enable machines to understand human conversation nuances.
* The firm uses a proprietary Ensemble Listening Model (ELM) architecture to generate models for the Velma platform.
* Velma integrates over 100 specialized models that detect emotion, tone, intent, synthetic speech, and conversational behavior.
* Modulate analyzes more than 10 million hours of audio each month, with over 600 million hours analyzed in total.
* Transcription and deepfake detection models ranked #1 on public benchmarks like Hugging Face.
* The platform focuses on analyzing voice rather than creating voice.
* Velma results are claimed to be twice as accurate as traditional LLMs and generate seven times fewer false positives.
* Applications include protecting healthcare, enhancing voice agent empathy, detecting extremism, monitoring voice agent performance, stopping child grooming conversations, and voice masking in high-risk scenarios.

The narrative centers on shifting the focus of AI security and intelligence from textual data to the modality of voice, positioning audio-native AI as a necessary defense against increasingly sophisticated manipulative uses of generative technology. The core tension lies between the rapid proliferation of voice deepfakes and the slow adaptation of existing AI frameworks that rely primarily on text. Modulate’s approach—building an ensemble of models focused on multi-faceted acoustic signals rather than simple transcription—suggests a necessary architectural shift toward multimodal, context-aware awareness in AI systems. The claim that audio-native AI can intervene "while it is in process" highlights a critical move from post-hoc detection to real-time systemic intervention, which has profound implications for agency in dynamic digital spaces. The investment strategy aims not just for technological superiority but for infrastructural enablement, suggesting the path forward involves democratizing this multi-layered intelligence layer so that application developers can focus on outcomes rather than rebuilding foundational perception systems. This raises questions about where the accountability for voice integrity resides when systems move from passive understanding to active intervention in real-time conversational contexts.
Bridge Questions: If voice becomes the primary interface, what ethical guardrails are necessary to prevent these detection and intervention capabilities from being weaponized against legitimate communication? How does the focus on detecting "synthetic speech" impact trust in all automated voice interactions across society? What infrastructure changes are required to support real-time, distributed acoustic intelligence at scale?