Introduction

Phonic, a small speech-to-speech startup, launched Phonic-1 today (October 29) alongside a $27.5M Series A led by First Round. It's a closed, API-only voice model: audio goes in, audio comes out, with no separate transcription or TTS step in between. You can try it right now through the Phonic API and agent builder, or with a one-line plugin in LiveKit Agents.

The headline is the benchmark: Artificial Analysis

on its Speech to Speech Index. But the feature I find genuinely new is that Phonic-1 can start a tool call while the caller is still talking, so there's no "let me look that up" pause at the end of your sentence.

What Shipped

From Phonic's launch post and docs:

  • Model: Phonic-1, model ID phonic-1, an end-to-end speech-to-speech model built for voice agents that need to think and use tools, not just chat.
  • Access: closed weights. Use it through the Phonic API, TypeScript and Python SDKs, a CLI, or the no-code agent builder. Phonic also lists self-hosted deployment for enterprise.
  • Price: from $0.07 per minute, one flat rate with no separate LLM or per-token charges.
  • Latency: Phonic quotes under 600ms from speech in to speech out.
  • Languages: trained on 50+ languages, and it can switch language mid-call while keeping the same voice.
  • Tool calls while listening: Phonic says it's the only speech-to-speech model that can act before the caller finishes speaking.
  • Small details that matter on phone calls: it's tuned for noisy lines, interruptions and turn-taking, and for reading out URLs, emails and alphanumeric codes correctly.

Phonic's own demo is a 44-minute call to a hotel in Brazil with 123 tool calls and two language switches, which you can listen to on the launch page.

The Numbers

Artificial Analysis tested Phonic-1 at its high intelligence setting through the API. Its

:

  • Speech to Speech Index: #1 of 22 at 87.1, a hair ahead of Grok Voice 3.0 at 86.9.
  • Tau Voice (agentic tasks): #1 at 79.9%, 5.4 points ahead of the next model.
  • Big Bench Audio (spoken reasoning): #3 at 98.9%, behind StepAudio 3 Realtime and Qwen Audio 3.0 Realtime Plus.
  • Time to first audio: 0.61s, the fastest of the top five models on the index. GPT-Live-1 and Gemini 3.8 Live Extended Thinking both land around 1.35s.
  • Speech Agent Arena: #8 of 30 on listener preference, #4 on task success at 92.3%.

Phonic also reports 83.2 on Sierra's multilingual full-duplex τ³-Voice test, versus 79.9 for GPT-Live-1 and 68.5 for Gemini 3.8 Live. That one is self-reported.

The honest bit I liked: in Phonic's own blind tests against real humans, people still preferred the human about 60% of the time. Phonic calls naturalness its main research focus for the next model.

How to Try It Today

The quickest path is to call the demo line on phonic.ai and talk to a Phonic agent. If you build with LiveKit, the plugin is live on day zero. In Python:

uv add "livekit-agents[phonic]~=1.8"
# then set PHONIC_API_KEY in your .env

from livekit.plugins import phonic

session = AgentSession(
    llm=phonic.realtime.RealtimeModel(voice="sabrina"),
)

No separate TTS plugin is needed, since the model speaks for itself, and session.say() works for scripted lines.

Pro Tip

Set min_words_to_interrupt=2 on the realtime model if you're testing in a noisy room or with a character that gets a lot of "mm-hm" from players. Single-word sounds then won't cut the agent off mid-line, which makes long conversations feel much less jumpy.

Why It Matters for Audio Creators

This is the second "smart speech-to-speech" launch in two days, after Rime Prism, and the pattern is clear: voice models are moving past the old chain of speech-to-text, then an LLM, then TTS. A single model hears the hesitation in your voice and answers in kind, which is exactly what you want for interactive characters, voiced game NPCs, AI hosts and audio stories that talk back.

Acting mid-sentence is the part to watch for creative work. A character that can check inventory, pull a quest detail or cue a sound effect while the player is still talking feels alive in a way that a pause-then-answer agent never does. The trade-offs are that it's closed, priced per minute and built for phone-style agents first, so it isn't a tool for rendering finished voiceover the way a TTS is.

Conclusion

Phonic-1 is a closed speech-to-speech model from Phonic that debuted at #1 on Artificial Analysis' Speech to Speech Index (87.1), answers in about 0.61 seconds, handles 50+ languages and can call tools while you're still speaking. It's live today on the Phonic API at $0.07 a minute and in LiveKit Agents. Call the demo line before you write a line of code; it's the fastest way to hear whether it fits your project.

—Anabel ♡