Voice Understanding AI Modulate Raises $25M

As voice becomes a new interface for AI, investors are betting not just on models that generate speech, but on models that understand it. Nuance Labs raised a $50 million Series A this month, and now Modulate has secured $25 million of its own.

Modulate positions itself as a general-purpose voice understanding engine, not tied to any single industry. Its Velma platform analyzes audio directly to detect signals like emotion, tone, intent, emphasis, synthetic speech, and conversational behavior — before any of it gets converted to text. Those signals combine to flag higher-level events such as fraud attempts, AI agent failures, harassment, customer dissatisfaction, and policy violations, and can trigger intervention while a conversation is still happening. The technology behind it is Modulate’s Ensemble Listening Model (ELM) architecture, which orchestrates more than 100 specialized audio models instead of relying on one massive foundation model, delivering up to 1,000x greater efficiency, according to the company. Modulate now analyzes more than 10 million hours of audio a month and has processed over 600 million hours in total. Its transcription technology ranks first on Hugging Face’s Open ASR Leaderboard, and its deepfake detection ranks first on the same platform’s deepfake speech benchmark, with 98.9% accuracy on public benchmark data.

The use cases go well beyond fraud. Modulate’s models protect healthcare institutions from deepfake attacks, help voice AI agents respond with more emotion and empathy, reduce extremism and harassment on social platforms, monitor voice agent performance, detect child grooming conversations, and mask the identity of agents operating in high-risk scenarios — fraud detection is just one line item on a much longer list.

Founded in 2017, Modulate was co-founded by Carter Huffman and Mike Pappas, who met at MIT after Pappas solved a physics problem Huffman was working out on a hallway whiteboard; the two prototyped startup ideas on the side while researching dark matter and cosmic expansion. CEO Huffman went on to NASA’s Jet Propulsion Laboratory after MIT, building machine learning models for spacecraft, and represented the US in a science research competition in high school for his work simulating dark matter candidate theories. Chairman Pappas previously worked on cloud security at Bridgewater Associates and now regularly discusses AI policy before the US Congress alongside representatives from OpenAI, Meta, and Microsoft. Around this round, Huffman moved into the CEO role while Pappas became chairman.

The new $25 million round was led by Future Ventures, with participation from Hyperplane and Lakestar — the latter a returning investor that also led Modulate’s $30 million Series A in 2022. TechCrunch reported the round brings Modulate’s total funding to $60 million. The capital will fund AI/ML research, product and engineering, developer relations, and partnerships. CEO Huffman said developers shouldn’t have to rebuild the audio understanding layer every time they build a new voice experience, and that this investment lets Modulate grow the team and move faster to provide that infrastructure. Future Ventures co-founder Steve Jurvetson said Modulate has already proven its technology in some of the most demanding voice environments, and that the market is now expanding quickly into AI agents, security, and customer experience.

The Competition

Looking only at narrow rivals misses the bigger threat: general-purpose models that swallow audio understanding whole. OpenAI‘s GPT-4o Realtime API can pick up on non-verbal cues like laughter and respond with five distinct emotions, and Google‘s Gemini native audio adjusts its responses to a user’s tone through a feature called Affective Dialog. Modulate’s own pitch — that 100+ specialized models beat one giant model by up to 1,000x on efficiency — is essentially an argument against exactly this kind of general-purpose approach.

Nuance Labs shares Modulate’s ambition to read and respond to vocal expression, building a foundation model that reads tone, gaze, and expression together; it raised $50 million this month. Hume AI, long known for emotion recognition, is often mentioned in the same breath, though its standing as an independent competitor is shakier since Google DeepMind poached founder Alan Cowen and his core engineering team in January.

In transcription and audio intelligence infrastructure, Deepgram is the most direct rival. It sells its own “Audio Intelligence” product bundling summarization, topic detection, sentiment, and intent analysis, and raised a $130 million Series C at a $1.3 billion valuation this January to become a unicorn — dwarfing Modulate in scale. Even so, Modulate claims to have taken the top transcription-accuracy spot on Hugging Face’s benchmark ahead of Deepgram, at roughly a seventh of the price (2.5–6 cents per hour versus Deepgram’s 31–55 cents).

Far more capital, in fact, is chasing voice generation than voice understanding. ElevenLabs surpassed $500 million in annual revenue and raised a $500 million Series D at an $11 billion valuation in February. It’s a different category, but the same broader wave — voice becoming AI’s next interface — that Modulate is riding.

In the narrower fraud and deepfake detection niche, Pindrop is the established vertical player, specializing in call-center voice authentication; it reportedly raised $100 million in venture debt in 2024.