The latest AI news and updates of Cartesia — Voice AI company building real-time speech models for live, human-like voice agents. Covering Cartesia's latest product updates, launches, and analysis from the past 90 days.
CartesiaSep 2Cartesia's Sonic-3.6 model ranks first across all eight customer support categories on Datapoint AI's newly released TTS benchmark. The ranking, based on 42,000 blind human votes, evaluates performance in areas including empathy, spell-outs, repairs, and disfluency. The benchmark uses the largest open-source human audio preferences dataset to measure model quality.
CartesiaAug 29Cartesia released Sonic-3.6, a text-to-speech model rebuilt from the architecture up to replace its previous version. The model achieves up to 93% listener preference in blind head-to-head tests across fifteen locales. It supports 44 languages, generates audio twice as fast as its predecessor, and delivers sub-90ms latency for real-time voice agent applications.
CartesiaAug 22Cartesia launched Sonic-3.6 in beta, featuring improved naturalness across 44 languages. The model ranks #1 on Artificial Analysis’ provider and controlled voice streaming leaderboards. It supports seamless Hinglish code-switching between Hindi and English, with fundamental model improvements derived from developer feedback.
CartesiaAug 15Cartesia released Sonic 3.6, enhancing naturalness and accuracy across nine existing Indic languages. The update refines Hindi flow and Hinglish code-switching while adding support for Urdu and Odia. These improvements provide more expressive and accurate voice generation tailored to India's diverse linguistic landscape.
CartesiaAug 13Cartesia integrated HeyGen’s LiveAvatar with its Line agent platform, enabling real-time, lip-synced video avatars for voice agents. The Cartesia Agent Connector bridges the agent’s audio output with the avatar’s visual rendering. This integration is available via the web app or API, allowing agents to maintain their existing prompt, tool access, and turn-taking logic.
CartesiaAug 11Cartesia updated its Ink-2 speech-to-text model with keyterm prompting and configurable turn detection. Keyterm prompting accepts rare domain-specific terms — product names, jargon, and foreign words — boosting keyword recall by 20% with zero added latency. Configurable turn detection enables endpointing tuned to prioritize either lower latency or higher accuracy for specific conversational flows.
CartesiaJul 28Cartesia published an analysis of text-to-speech evaluation challenges, identifying five dimensions of quality: correctness, naturalness, contextual correctness, robustness, and audio quality. The post argues that most current benchmarks rely on Word Error Rate, which only measures correctness, failing to capture the nuanced prosody, streaming stability, and domain-specific accuracy required for enterprise-grade voice agents.
CartesiaJul 16Cartesia integrated Parallel Search’s Turbo mode into its Line voice agent platform. Agents now query the live web at conversational speed without interrupting the natural flow of speech. The integration uses a loopback tool to fetch current information, keeping search payloads compact for live voice turns.
CartesiaJul 9Cartesia launched Ink-2, a speech-to-text model built for real-time voice agents. It features semantic endpointing to manage turn-taking without external voice activity detection and achieves a 0.1s time-to-final-transcript latency. The model ranks first on Artificial Analysis’s streaming leaderboard for final transcript accuracy, with a 3.59% word error rate. It is available via API and integrated platforms.
Frequently asked questions