The Unicorn Round
Voice AI has been "almost ready" for two decades. The demo works; the drive-thru misunderstands the order; the call-center bot traps you in a loop. Then a company that says it didn't need capital raises $130 million at a $1.3 billion valuation anyway, TechCrunch reported, and ships a unified multilingual model that runs on-device.
The Series C closed in January. AVP led; partner Elizabeth de Saint-Aignan said her team kept hearing voice AI come up in enterprise conversations throughout 2024 (contact centers, sales development, drive-thru lanes) and traced a surprising share of deployments back to Deepgram. Existing backers Alkeon, In-Q-Tel, Madrona, Tiger Global, Wing, and Y Combinator doubled down. New names joined: Alumni Ventures, Columbia University, Princeville Capital, Twilio, and SAP. Total raised now pushes past $215 million, according to TechCrunch.
CEO Scott Stephenson told TechCrunch the company ran cash-flow positive in 2025 and wasn't shopping the deal. "We had multiple people coming to us," he said. "We wanted strategic investors who understand voice AI and the technical intricacies and have relationships with companies building on the technology." That selectivity shows in the list: Twilio and SAP are platform players embedding voice into broader stacks; In-Q-Tel signals defense traction.
The unicorn badge lands amid a cluster of outsized voice AI rounds: Sesame at $250 million, TechCrunch's figures show, ElevenLabs at $180 million, Gradium at $70 million. Analysts project the voice market growing above 30 percent year-over-year to $14–20 billion by 2030, as market.us and Grand View Research report. At that rate, model and API providers can become multi-billion-dollar infrastructure layers if they solve the latency, accuracy, and multilingual gaps that still make most voice interactions feel brittle.
Deepgram counts more than 1,300 organizations on its APIs, including meeting notetaker Granola, voice agent startup Vapi, and Twilio itself. The company also acquired Y Combinator–backed OfOne, whose drive-thru system hits 93 percent order accuracy. Stephenson framed the bet bluntly: "Food ordering might be the first positive interaction more than 300 million Americans have with voice AI. There have been a lot of sour interactions over the last 20 years… But when you can order your food using natural conversation, people would think the technology is ready."
The valuation signals that investors are betting the voice layer is finally shifting from experiment to infrastructure, and that the company powering Granola, Vapi, and a growing slice of drive-thru lanes is positioned to own a meaningful share of it.
Moving Voice Off the Cloud
Deepgram's technical push since the funding round centers on two launches that move conversational voice AI off the cloud and onto devices: Flux Multilingual, a unified speech-to-text model handling ten languages in a single conversation, and a Snapdragon deployment running Nova-3 recognition locally on Qualcomm's PC platform. Together they address the latency-interruption tradeoff that has kept voice agents from feeling natural, without sending audio to a central server.
Flux Multilingual supports English, Spanish, German, French, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch in one model. The company calls it the first conversational speech recognition model: it detects turn boundaries, handles barge-in interruptions without cutting off the speaker, and maintains low latency end-to-end. Stephenson said the team "pushed up the launch of Flux so that we could launch it here at Vapycon," a signal that the conversational AI developer community was the primary audience. The model runs through a single Voice Agent API unifying speech-to-text, text-to-speech, and LLM orchestration, eliminating the stitching of separate components that adds latency and failure points in cascaded architectures.
Deepgram's quality metric, VAQI — Voice Agent Quality Index — folds latency, interruption rate, response coverage, and other signals into one number. Stephenson framed the design priority bluntly: "We focused on accuracy first." Word error rates have fallen from roughly 30 percent in the early transformer days to around 5 percent on general audio and as low as 1 percent on adapted models, a trajectory driven by training on tens of millions of hours rather than thousands. Flux English and Flux Multilingual ship with built-in turn detection and natural interruption handling; Nova-3 Monolingual and Nova-3 Multilingual remain the highest-accuracy options for batch and far-field scenarios with noise, crosstalk, and multiple languages.
The SDK (v7.6.0 as of July 2026) plugs into jambones, LiveKit, Vapy, Pipecat, and Cloudflare Workers, with testing tooling via Coval. Early customers stress-tested the stack during a free October tier allowing up to 50 concurrent connections. Pricing afterward: $0.075 per minute for the standard Voice Agent API tier, $0.163 per minute for advanced; Aura-2 text-to-speech runs $0.030 per 1,000 characters. Compliance certifications (SOC 2 Type 1 and 2, HIPAA, GDPR with EU data residency, CCPA, PCI) bake into the platform, not bolt on.
The Snapdragon integration brings Nova-3 to Qualcomm's neural processing units on Windows-on-Snapdragon PCs, meaning the model runs locally without a cloud round-trip. That matters for defense and industrial settings where connectivity is intermittent or data cannot leave the device. The architecture mirrors Neuroplex: an end-to-end approach preserving acoustic context (prosody, hesitation, overlapping speech) instead of collapsing everything to text for an LLM and then re-expanding to audio. Stephenson called the cascaded alternative an "everything-down-into-text" pipeline that loses richness at each stage. Flux is the first piece of Neuroplex; a forthcoming text-to-speech announcement will extend the same architecture to the output side.
Stephenson's timeline: clear an "audio Turing test" — five minutes of conversation indistinguishable from a human — in B2B use cases within a year, in arbitrary use cases within two. The Snapdragon deployment and Flux's multilingual, interruption-aware design are the concrete steps toward that mark.
Where the Edge Matters
Deepgram's edge deployment on Snapdragon chips and the Flux Multilingual model target environments where cloud connectivity is unreliable, latency is a safety issue, and data cannot leave the device. The architecture meets defense requirements for air-gapped or classified networks; the company's self-hosted and on-prem options, coupled with certifications baked in, hand procurement offices a checklist they can actually tick.
Multilingual support shifts the calculus for global operations. Flux ships with ten languages; the broader Nova family covers dozens more. A single model that switches languages mid-conversation without reloading matters when a multinational crew works a shared production line or when allied forces share a maintenance depot. The CEO said the architecture injects context and emits end-of-turn signals, reducing the false interruptions that plague voice agents in loud environments.
Enterprise customers embed the stack. Sierra, Kore.ai, and NiCE Cognigy cite Deepgram APIs as their voice layer, praising reliability and low latency at scale. The company's "Deep enterprise runtime" distributes that stack across a global partner network, which Stephenson called the infrastructure to ship "a billion interactions that happen every day" — not a science experiment. Pricing pressure from hyperscalers exists, but Deepgram set the floor: a 10× price drop that unlocked 1,000× usage, the founder said. That floor now extends to the edge.
The next step is Neuroplex, a forthcoming orchestration layer Stephenson previewed at that developer conference, designed to stitch ASR, TTS, and LLM calls into a single runtime partners can embed. Deepgram accelerated Flux to launch there; the company made it free for October to seed adoption. For defense integrators and robotics OEMs, the signal stands clear: the voice layer is no longer a cloud service you bolt on. It is a model you compile into firmware, certify once, and deploy on hardware that already sits in the rack.
The Race Compresses
ElevenLabs entered the STT market in early 2025 with Scribe v1 and has since released Scribe v2 Realtime, which hits 150 milliseconds end-to-end across dozens of languages at better than 93 percent accuracy on the FLEURS benchmark. The company followed it with Slam-1, a speech-language model released late last year that blends transcription and understanding in a single pass, and added its own TTS stack, giving developers a single-vendor voice loop that mirrors the full-stack pitch Deepgram bundles at roughly $4.50 an hour.
NVIDIA took a different lever. Its Canary Qwen 2.5B model — a FastConformer encoder paired with a Qwen3-1.7B LLM decoder — sits atop the Hugging Face Open ASR Leaderboard with a 5.63 percent word error rate, edging past Deepgram's self-reported 5.26 percent on Nova-3 batch. Because Canary is open-weight, it gives robotics and defense integrators a path to on-prem deployment without an enterprise contract, a friction point all three major API vendors still route through sales-led deals. NVIDIA's model also runs efficiently on Jetson Orin and IGX platforms, making it a natural rival for the Snapdragon edge story Deepgram just activated.
Speechmatics, the longest-standing pure-play in the group, has responded by emphasizing breadth over latency bragging rights. Its streaming tier covers 55-plus languages and supports more deployment topologies (cloud, VPC, air-gapped appliance) than either Deepgram or AssemblyAI. AssemblyAI's Universal-3 Pro Streaming, by contrast, supports only six real-time languages but bundles transcript intelligence (summarization, PII redaction, topic detection) that Deepgram prices separately or leaves to the LLM layer.
The hyperscalers have not stood still. Google Cloud now offers 125-plus languages across Chirp 2, Chirp 3, and specialized medical and telephony variants. AWS Transcribe handles 100-plus languages with automatic language identification, while Azure Cognitive Services offers Custom Speech for fine-tuning on proprietary vocabulary. For teams already locked into those clouds, the switching cost remains high enough that Deepgram's pricing advantage has to clear a procurement threshold, not just a benchmark.
| Provider | Batch (per min) | Streaming (per min) |
|---|---|---|
| Deepgram | $0.0043 | $0.0077 |
| Google Standard | $0.016 | — |
Independent benchmarks have become the arena where these claims collide. Procurement memos now cite Artificial Analysis and the Hugging Face Open ASR Leaderboard more often than vendor white papers, because vendor numbers "don't reproduce consistently across different test sets," as Deepgram's own comparison page acknowledges. Deepgram measured Nova-3's claimed improvement over its nearest competitor on its internal nine-domain suite; on noisy, accented, code-switched audio the gap narrows sharply: a 95 percent system on clean English drops to 85 percent on accented speech with background noise, the Future AGI 2026 guide notes.
The market has shifted from "who has the lowest WER on LibriSpeech" to "who delivers sub-300-millisecond STT at scale on messy audio with multilingual code-switching and a deployment model that clears security review." Deepgram leads on latency and cost-efficiency; ElevenLabs leads on multilingual accuracy at real-time speeds; Speechmatics leads on deployment flexibility and language count; NVIDIA leads on open-weight quality; the hyperscalers lead on ecosystem lock-in. The new funding gives Deepgram runway to defend its latency edge and fund the Snapdragon optimization work, but the competitive half-life of any lead in this segment now measures in quarters, not years.
Working in frontier tech? Zero G Talent tracks the openings: see every open Leap role, browse frontier tech jobs, the companies hiring, and the people building the field.