Phonic released its Phonic-1 speech-to-speech model on October 29. The company says the model can begin tool calls while someone is still speaking, and retain vocal cues that separate transcription and speech-generation stages can lose. Its launch includes a 44-minute hotel-call demonstration with 123 tool calls, alongside examples of switching languages within a conversation.
Phonic, a company developing speech-to-speech systems, released its Phonic-1 model on October 29. It takes spoken input and generates spoken output. Conventional voice-agent cascades use separate models to transcribe speech, determine a response and synthesize a voice. Phonic says its approach aims to preserve cues such as tone, pauses and the way a person pronounces a name. The company has focused on this architecture for more than two years.[1]
The company says Phonic-1 can start tool calls before the caller finishes speaking. Its travel-request demonstration shows several calls beginning during a request. Phonic’s post-training targets conversational flow, turn-taking in noisy environments and sensitivity to a speaker’s delivery. Phonic says a speaker’s identity can stay consistent when the conversation moves between languages. These capabilities are presented through the company’s launch demonstrations.[1]
In another demonstration, a caller plans a bachelorette party with a Brazilian hotel. The 44-minute example contains 123 tool calls. The platform makes the call trace available, showing its written transcript, delays between turns and each invoked tool. Phonic names customer support, healthcare, recruiting and logistics among its application areas. Ashby, a recruiting-software customer, says it uses Phonic for artificial intelligence interviews in 25 countries. The model is accessible through Phonic’s platform.[1]
