🌏 中文版
Funding Details
| Item | Value |
|---|---|
| Company | Phonic (San Francisco, US) |
| Round | Series A |
| Amount | $27.5M |
| Lead | First Round Capital |
| Participants | Lux Capital, Felicis, Definition, Essence VC |
| Valuation | Undisclosed |
| Total raised | $31.5M ($4M seed + $27.5M Series A) |
| Founded | 2024 (the team describes founding the company "more than two years ago") |
| Headcount | About 10 (researchers and engineers from MIT, Stanford, and CMU) |
What the company does
Phonic builds the underlying model layer for voice agents. Instead of stacking another layer on top of the typical speech-recognition (STT) + LLM + text-to-speech (TTS) pipeline, it trains a native, end-to-end speech-to-speech model so a voice agent can hear tone, pauses, and emotion directly — not just the semantic meaning left after converting speech to text.
Its differentiator is acting while the caller is still talking. Most voice agents wait for the user to finish a sentence and say "let me check" before calling a tool; Phonic-1 can call tools while the caller is mid-sentence, so the answer is ready the moment they stop. The company has demonstrated a 44-minute call with 123 tool calls that switches seamlessly between English and Portuguese. The model supports more than 50 languages and can switch languages mid-call while keeping the same voice identity. Customers include Ashby (powering AI interview scheduling in 25 countries), Maven AGI, Rho, and a Fortune 50 IT infrastructure company, running on Together AI's infrastructure.
Phonic-1 (High) debuted at number one on Artificial Analysis's Speech-to-Speech Index (87.1, narrowly ahead of Grok Voice 3.0's 86.9) and has the fastest time-to-first-audio among the top five models (0.61 seconds). On Sierra's multilingual full-duplex benchmark, τ³-Voice, it scores 83.2, ahead of GPT-Live-1's 79.9 and Gemini 3.8 Live's 68.5.
What this round signals
What it means for the agent ecosystem
The bottleneck for voice agents is shifting from "how do we stitch three models together" to "should we train a native end-to-end model instead." Phonic's "act while listening" pitch directly targets the two long-standing problems with cascaded pipelines in real customer-service calls: latency (waiting for the caller to finish before processing starts) and lost emotion (converting speech to text strips out tone, pauses, and feeling). That suggests the next round of voice-agent competition will likely be decided at the model layer rather than the orchestration layer — tool calling and memory management.
What investors are betting on
First Round led the round, but the more notable detail is the roster of developer-tool angels who participated personally: Vercel's Guillermo Rauch, Replit's Amjad Masad, Hugging Face's Clement Delangue, Modal's Erik Bernhardsson, and AngelList's Naval Ravikant. That roster reads less like a group of financial investors betting on market size, and more like the developer-infrastructure ecosystem collectively betting that voice will become the next agent interface.
Numbers worth watching
Figures from Phonic's official blog (phonic.ai/blog/phonic-1, 2026-10-29), from blind pairwise comparisons run by Podonos. Value axis 0-100%, x=200+value x 4. The filled blue point always represents Phonic-1's score; the hollow gray point represents the comparison model or human score.
- On the multilingual full-duplex benchmark τ³-Voice, Phonic-1 scores 83.2 — 3.3 points ahead of GPT-Live-1 (79.9) and 14.7 points ahead of Gemini 3.8 Live (68.5)
- It has the fastest time-to-first-audio among the top five models at 0.61 seconds — one of the rare cases that's best on both intelligence and latency at once
- Just $31.5M in total funding produced a model that debuted at number one, with a team of only about 10 people — notably capital-efficient in a wave of voice AI funding where rounds routinely run into the hundreds of millions
Watchlist status
Phonic isn't on the watchlist yet. Recommend adding it under section C5 (Voice AI / Voice Agent) — the existing names there (Vapi, Retell, Bland AI, ElevenLabs, PolyAI, Parloa, Synthflow, Cartesia, LiveKit) are all cascaded or platform-layer solutions, while Phonic's differentiation is the native speech-to-speech model itself. Tracking: tool-calling while the caller is still speaking, $27.5M Series A led by First Round.
Today's takeaway
Phonic publicly disclosed that it loses to human voice actors in blind tests (32.0% vs. 57.0%), rather than only showing off the three AI models it beat. That kind of disclosure — being upfront about exactly how much you're still losing by — is more trustworthy than pure marketing copy. It's also a reminder that the real competitive bar for voice agents today isn't "beat other AI models," it's "close the gap with humans" — and that gap is still 25 percentage points wide.