Phonic-1 starts tool calls before the caller stops talking
A new speech-to-speech model replies 1.1 seconds after a caller finishes, against 15.7 seconds or more for a typical cascaded voice pipeline.
Phonic released Phonic-1 on 29 October, a speech-to-speech model built for voice agents that need to act while a caller is still talking, not just after they finish. Rather than chaining a transcription model, an LLM and a text-to-speech model together, it processes audio directly and starts executing tool calls mid-sentence: in Phonic’s own comparison, it replies 1.1 seconds after the caller stops speaking, against 15.7 seconds or more for a cascaded pipeline doing the same task.
The model is trained on more than 50 languages and can switch between them mid-conversation while keeping the same voice. It ranks first on Artificial Analysis’s third-party Speech-to-Speech Index, and scores 83.2 on Sierra’s multilingual tau-cubed-Voice benchmark, against 79.9 for OpenAI’s GPT-Live-1 and 68.5 for Google’s Gemini 3.8 Live. Pricing starts at $0.07 a minute with no minimum commitment, or $0.065 a minute on a plan with a $200 monthly minimum.
Why it matters: the 1.1-second figure is the real headline. Voice agents that wait until a caller finishes speaking before calling any API feel laggy, and that lag is usually the first complaint about automated phone support. A model that starts the tool call before the sentence ends closes that gap without you having to re-architect your own pipeline.
The caveat: the benchmark numbers come from Phonic’s own announcement, including a third-party index it did not run itself, so they are worth confirming against your own call flows before switching.