Phonic and Rime Unveil Voice AI Models

Phonic announced Phonic-1 on October 29, 2026, a voice AI model that can run external tools while a user is speaking. Rime announced Prism the previous day, October 28: it lets companies inspect an AI response as text before the system speaks it. Both models directly process users’ audio and respond with audio, a Speech-to-Speech approach. Phonic-1 emphasizes taking action during a user’s speech, while Prism gives companies control over what the system says.

Voice AI can also use a cascade architecture: speech recognition turns spoken words into text, a language model creates a response, and speech synthesis reads it aloud. Although these steps are linked in sequence, a transcript alone can miss vocal tone, hesitation, and pauses. The companies said their models work directly with audio to use such cues in conversation, and also demonstrated support for tasks that use external systems.

Phonic-1 can call tools and begin processing before a user finishes speaking. In a travel-request comparison demo published by Phonic, it called five tools while the user was speaking and replied 1.1 seconds after the user finished. GPT-Live called tools only after the user finished and took 15.7 seconds to respond. These figures are results from the company’s specific demo. Phonic-1 was trained in more than 50 languages and can keep the same voice when the conversation changes languages. Phonic also published a demo switching from English to Spanish, Japanese, and Chinese, and another showing 123 tool calls during a 44-minute conversation.

Phonic-1 was available on its announcement day through an API, TypeScript and Python SDKs, and a no-code agent-building feature. Phonic named customer service, job interviews, and patient intake as intended uses.

Prism was trained as an integrated system that listens to users, generates a text response, and then speaks it. It continues listening while the AI is speaking and can run tools at the same time. The response text provides a checkpoint where a company’s application can check the content against internal rules and restrictions before users hear it. Prism directly processes audio, including cues such as hesitation, anger, and laughter, while giving companies control over responses.

Rime evaluated tool use and instruction following on τ-Voice, a benchmark modeled on customer service in aviation, retail, and telecommunications. Rime reported a score of 56.6 without adding extra reasoning before speech, and 0.57 seconds from the end of a user’s utterance to the first audio output. Prism supports fine-tuning on company call data and deployment on-premises, in a virtual private cloud (VPC), or in the cloud. At announcement, APIs for reinforcement learning (RL) and supervised fine-tuning (SFT), which companies can use for additional training, were being developed with beta customers. The official documentation lists English as the only supported language.

Phonic announced Phonic-1 on October 29, 2026, a voice AI model that can run external tools while a user is speaking. Rime announced Prism the previous day, October 28: it lets companies inspect an AI response as text before the system speaks it. Both models directly process users’ audio and respond with audio, a Speech-to-Speech approach. Phonic-1 emphasizes taking action during a user’s speech, while Prism gives companies control over what the system says.

Voice AI can also use a cascade architecture: speech recognition turns spoken words into text, a language model creates a response, and speech synthesis reads it aloud. Although these steps are linked in sequence, a transcript alone can miss vocal tone, hesitation, and pauses. The companies said their models work directly with audio to use such cues in conversation, and also demonstrated support for tasks that use external systems.

Phonic-1 can call tools and begin processing before a user finishes speaking. In a travel-request comparison demo published by Phonic, it called five tools while the user was speaking and replied 1.1 seconds after the user finished. GPT-Live called tools only after the user finished and took 15.7 seconds to respond. These figures are results from the company’s specific demo. Phonic-1 was trained in more than 50 languages and can keep the same voice when the conversation changes languages. Phonic also published a demo switching from English to Spanish, Japanese, and Chinese, and another showing 123 tool calls during a 44-minute conversation.

Phonic-1 was available on its announcement day through an API, TypeScript and Python SDKs, and a no-code agent-building feature. Phonic named customer service, job interviews, and patient intake as intended uses.

Prism was trained as an integrated system that listens to users, generates a text response, and then speaks it. It continues listening while the AI is speaking and can run tools at the same time. The response text provides a checkpoint where a company’s application can check the content against internal rules and restrictions before users hear it. Prism directly processes audio, including cues such as hesitation, anger, and laughter, while giving companies control over responses.

Rime evaluated tool use and instruction following on τ-Voice, a benchmark modeled on customer service in aviation, retail, and telecommunications. Rime reported a score of 56.6 without adding extra reasoning before speech, and 0.57 seconds from the end of a user’s utterance to the first audio output. Prism supports fine-tuning on company call data and deployment on-premises, in a virtual private cloud (VPC), or in the cloud. At announcement, APIs for reinforcement learning (RL) and supervised fine-tuning (SFT), which companies can use for additional training, were being developed with beta customers. The official documentation lists English as the only supported language.