August 2026 Summaries

Cartesia announced Sonic-3.6, a multilingual text-to-speech model that it says improves substantially on Sonic-3.5 in naturalness, voice quality, emotional delivery, and context-aware intonation. The company reports that listeners preferred the model in up to 93% of blind comparisons across 15 locales, including a 92% preference over Eleven v3 for U.S. English, and says it ranks first on Artificial Analysis leaderboards. Sonic-3.6 supports 44 languages, 61 locales, more than 500 preset voices, and newly added Odia and Urdu, while offering features such as Hindi-English code-switching, locale-specific pronunciation of dates and numbers, and improved accent retention in instant voice clones. Designed for enterprise deployment, it reportedly responds in under 90 milliseconds, generates speech nearly twice as fast as v3 Conversational, supports cloud, on-premises, and localized deployments, and includes a 99.9% uptime SLA and compliance coverage for SOC 2, PCI, HIPAA, and GDPR. The model is generally available through Cartesia’s playground and API.

Cartesia has introduced keyterm prompting and configurable turn detection for its Ink-2 English streaming speech-to-text model. Keyterm prompting lets users provide up to 100 domain-specific terms, such as brand names, drug names, account details, and industry jargon, to improve transcription accuracy without added latency; the company reports a 20% recall improvement on its Earnings22 benchmark and fewer missed keywords than Deepgram Flux. Cartesia also reports gains on its enterprise-focused EVA-Bench across domains including airlines, IT services, and HR. Configurable turn detection allows developers to prioritize lower latency by ending turns sooner or greater endpointing accuracy by waiting through pauses, with settings available both when opening a connection and during a live stream.