Rime, a San Francisco voice AI startup, closed a $24 million Series A on July 15, led by M13 Ventures with participation from Corazon Capital, Unusual Ventures, Cadenza Ventures, and Twilio Ventures. The round follows a $5.5 million seed last May, bringing total funding to roughly $30 million. Paris-based Gradium, which launched out of stealth in December 2025 with $70 million from FirstMark Capital and Eric Schmidt, reopened its seed round to bring total funding to $100 million with new investment from Nvidia. Combined, the two companies have raised $124 million in a sector that barely registers in most AI coverage cycles.
The capital is flowing into voice AI at a moment when coding agents, chat interfaces, and autonomous browsing dominate the industry’s attention. Yet the investment thesis is straightforward: enterprise phone systems still handle billions of calls per year, and most of them cannot be replaced by chat widgets or email bots. Healthcare intake calls, banking fraud alerts, hotel booking confirmations, insurance claims triage. These workflows run on phones because the customers on the other end expect phones. Voice AI startups are betting that the fastest path to enterprise agent adoption runs through the interface enterprises already use.
The Rime Approach: Linguistics Over Scale
Rime’s strategy is unusual in an industry that defaults to scale. Rather than training on scraped web audio, the company built a recording studio in San Francisco where it captures conversational data directly. The goal is phoneme-level control over pronunciation, accent, and rhythm. According to Artiverse, Rime CEO Lily Clifford described the current state of voice AI as “an improved version of old IVR systems,” with better-sounding voices but limited conversational engagement.
The company offers over 600 voices across 50 languages and reports powering more than 1.5 million minutes of enterprise conversation across fintech, healthcare, and hospitality. Sub-100-millisecond time-to-first-byte latency is the technical benchmark the company promotes. Ali Mansoor, founding engineer at Trillet AI, told Rime that their co-located endpoint delivers “a 3x latency improvement” over alternatives, calling sub-100ms TTFB “the difference between a conversation that feels human and one that doesn’t.”
The Series A also brought a significant hire. Rafael Valle, formerly leading audio understanding at Meta’s Superintelligence Labs and previously part of Nvidia’s Applied Deep Learning Research team, joined Rime as Chief Scientist. Valle’s background spans the two companies (Nvidia and Meta) that have done the most to advance audio and speech model infrastructure, and his move to a 30-person startup signals where he sees the frontier moving.
M13 Partner Morgan Blumberg, who joined Rime’s board, framed the investment around infrastructure timing: “The design patterns that will define the speech-to-speech era of AI haven’t been built yet. Rime is doing the foundational work by combining frontier AI with deep linguistic expertise to create the speech infrastructure the next generation of AI products will rely on.”
Gradium: The French Lab With Nvidia’s Backing
Gradium takes a different approach. Co-founded by Neil Zeghidour, who previously worked at Google Brain, DeepMind, and Facebook, the company operates out of Paris and grew from the French AI lab Kyutai. Its $100 million in total funding makes it one of the best-capitalized voice AI startups globally, and Nvidia’s direct participation in the round creates a hardware-software alignment that competitors lack.
Where Rime focuses on enterprise phone calls with phoneme-level control, Gradium has built a broader product surface. According to its product page, the company offers text-to-speech, speech-to-text, speech-to-speech translation, live translation, on-device TTS, and voice cloning. The breadth suggests Gradium is positioning itself as a full-stack voice infrastructure provider rather than an enterprise calling specialist.
The Nvidia connection matters beyond the capital. Voice AI inference at enterprise scale requires significant GPU compute for real-time processing, and a direct relationship with Nvidia provides both hardware access and potential co-optimization of inference pipelines. Gradium’s Bay Area office expansion, announced alongside the funding, suggests a push to win U.S. enterprise customers while leveraging French AI research talent.
Why Voice Attracts Defensible Capital
The voice AI funding pattern diverges from the broader venture market in a telling way. While AI wrapper companies struggle to raise capital amid concerns about moat erosion, voice startups are pulling in nine-figure rounds. The reason maps directly to what Tiger 21 founder Michael Sonnenfeldt told Forbes about investor risk calculus in July 2026: “The AI landscape has changed so radically that if you started from scratch, you could accomplish something that would be similar, but at a dramatically lower cost.”
Sonnenfeldt’s observation applies to chat-based AI products, where switching costs are negligible and model improvements can vaporize an entire product category overnight. Voice AI companies face a different competitive dynamic. Rime’s proprietary recording studio, its phoneme-level pronunciation system, and its sub-100ms latency are not capabilities that a competitor can replicate by calling a better foundation model. These are infrastructure advantages that compound over time: more recorded conversational data produces better voice models, which attract more enterprise customers, which generate more conversational data.
Twilio Ventures’ participation in Rime’s round adds another dimension. Twilio operates one of the largest communications infrastructure platforms in the world, handling billions of phone interactions annually. Its investment in Rime is not speculative: it represents a bet that Twilio’s own customers will need AI-powered voice agents integrated into existing telephony infrastructure. The distribution channel is already built. The missing piece is a voice model that sounds good enough to keep enterprise callers on the line.
The Foundation Model Tailwind
Voice AI’s capital surge coincides with a specific technical development at the foundation model layer. According to Towards AI’s weekly briefing, OpenAI shipped GPT-Realtime-2.1 and a mini variant in the same week, “improving interruption handling, noisy audio, and alphanumeric recognition while cutting p95 latency by at least 25%.” The improvement targets exactly the problems that enterprise voice agents face: background noise on customer calls, interruptions from confused callers, and the need to accurately capture account numbers, dates, and spelled-out names.
Foundation model providers are not competing with Rime and Gradium. They are improving the reasoning layer that voice AI companies build on top of. As frontier models get better at understanding intent in messy real-world audio, the value of specialized voice infrastructure increases rather than decreases. The model handles reasoning. The voice layer handles how it sounds, how fast it responds, how it pronounces the customer’s medication name, and whether it can run inside a HIPAA-compliant environment.
This is the structural distinction that makes voice AI defensible where chat wrappers are not. A chat-based AI product competes directly with the foundation model’s own chat interface. A voice AI product occupies a different layer: it converts the model’s intelligence into a medium that legacy enterprise systems already speak. Chat is a new interface that requires customer adoption. Voice is the existing interface that requires engineering.
The Market Nobody Counted
The scale of enterprise voice traffic is rarely cited in AI market sizing because it does not fit neatly into the “agents” category that dominates industry reports. But it is enormous. Contact centers alone handle over 100 billion calls globally per year. Healthcare systems route millions of patient intake calls daily. Hotels, airlines, insurance companies, and financial services firms all maintain phone systems because their customers demand it.
Most of these calls follow predictable patterns: verify identity, gather information, route to the right department, confirm a booking, explain a billing charge. These are exactly the workflows that AI agents can automate, but only if the agent can speak. A chatbot that handles insurance claims cannot replace a phone system that handles insurance claims. The medium is the constraint.
Rime’s Fortune 500 customer base and Gradium’s $100 million war chest suggest that enterprise buyers are already making this calculation. The question is not whether voice agents will handle enterprise calls. The question is which voice infrastructure layer enterprise buyers will standardize on.
The Consolidation Clock
Voice AI is still early enough that both Rime and Gradium can grow without direct competition. Rime’s focus on enterprise phone conversations with phoneme-level control targets regulated industries where pronunciation accuracy and latency are non-negotiable. Gradium’s broader product surface, spanning TTS, STT, translation, and voice cloning, positions it as a platform play.
But the market will not stay fragmented. As foundation models continue improving their real-time audio capabilities, the specialized voice layer will face pressure from above (OpenAI, Google, and Meta building voice directly into their APIs) and from below (enterprises building internal voice pipelines on open-source models). The $124 million in fresh capital buys Rime and Gradium a window to establish the kind of enterprise integration depth and switching costs that make displacement expensive.
Lily Clifford’s framing in Rime’s announcement captures the bet: “For the first time in human history, a rich understanding of language is available as a utility, delivered via API. Voice will be the primary way we interact with intelligence.” If she is right, the companies building that voice layer are not niche AI startups. They are the next generation of enterprise communications infrastructure.