Phonic launches a speech-to-speech model built for interruptions and language switches

Phonic launched a speech-to-speech model designed for interruptions and language switching.

Phonic announced Phonic-1 on October 29 alongside a $27.5 million Series A. The company says the voice model is available and describes it as processing and generating speech directly. It is designed to support full-duplex conversations, where a system can listen while it is speaking, and to switch languages within a conversation. [Phonic announcement](https://phonic.ai/blog/phonic-1)

Why timing changes the experience of a call

In a conventional pipeline, recognition converts the caller's audio to text, a model decides what to say and a synthesizer turns that response into speech. A speech-to-speech model can use acoustic information that a transcript may omit, such as whether a pause sounds like the end of an answer. The surrounding application still has to decide what to do when a caller interrupts or when network delays make the order of events unclear.

For example, a caller might begin correcting a date before the assistant finishes reading it back. A useful test checks whether playback stops, whether the corrected date is captured and whether the assistant confirms the updated value before booking anything. A first-audio measurement covers only one part of that interaction.

What the published measurements show

Phonic says the model was trained on more than 50 languages. It reports a score of 83.2 on Sierra's multilingual full-duplex voice benchmark and cites third-party listener comparisons for perceived naturalness. In the comparison chart involving human voice actors, listeners preferred the human responses more often. These results cover the tasks and samples evaluated. They do not give a per-language quality guarantee or establish how the model will perform with a particular telephony provider. [Phonic's test results and comparison methodology](https://phonic.ai/blog/phonic-1)

The launch offers voice teams a different architecture to evaluate. A deployment test can track interruption recovery, language changes, tool-call accuracy and overall task completion in the same recorded scenarios. That would show whether a voice that responds quickly also handles the decisions the call is meant to complete.