Sonic-3

Cartesia's real-time speech synthesis model, officially claimed as the first with an audio package below 90 milliseconds, a single model covering 44 languages, supporting voice cloning and custom pronunciation dictionaries, is mainly targeted at the voice Agent scenario.

AI Voice Tools#AI Voice#Speech Generation#Audio Processing

Updated: 2026-10-27

What is Sonic-3?

Sonic-3 is Cartesia's real-time text-to-speech (TTS) model, and its official page currently promotes Sonic-3.6. It is not targeted at dubbing export but at voice agents: it requires almost no waiting during conversations while pronouncing tones, numbers, and proper nouns correctly. For teams building voice products, the priority order for choosing a TTS is usually latency, followed by stability, and finally, how pleasant the voice sounds. The official page also offers two other product lines: speech recognition and hosted agents, making it easy to integrate recognition, synthesis, and conversation.

Who is it for?

Teams building voice customer service and call centers: Require low latency and stable call handling.
Voice Agent developers: Integrate real-time TTS as a layer in the dialogue chain into their products.
Overseas product teams: One system needs to cover multilingual broadcasting.
Teams for audio content and brand broadcasting: Need a unified timbre and accurate pronunciation.
Logistics and order-related businesses: Need to accurately read out order numbers, amounts, and addresses.

Key features

Low Latency

The official states that the first audio packet is returned within 90 milliseconds, which is a key metric for real-time conversation capability; end-to-end latency also includes speech recognition and network time.

Single model supporting 44 languages

The official states that one model covers 44 languages, facilitating the launch of the same system for multilingual use. There is no need to maintain separate models for each language, and the same voice can be reused across languages.

Expressiveness

The official description states that the voice can laugh at the right moments and accurately read out numeric strings such as order numbers, making it suitable for customer service and announcement scenarios; the tone and rhythm are adjusted according to the context, and the official goal is to make conversations sound like a human speaking.

Voice Cloning

The official lists voice cloning as an independent capability, allowing the replication of timbres for unified brand announcements.

Custom Pronunciation Dictionary

You can set the pronunciation for your own personal names, brand names and abbreviations to avoid misreading them every time, and also to maintain consistency in mixed Chinese-English text; the dictionary is maintained on a project basis, and you can set different pronunciations for the same brand name for different languages.

Deployment and Security

The official states that enterprise-grade deployment options from cloud to on-premises are available, and additionally, the Ink speech recognition and hosted Agent products can be integrated.

How to use Sonic-3

First, use the real text from your business for a trial listen, focusing on numbers, amounts, abbreviations, and mixed Chinese and English, then compare the stability of the same sentence with three different tones.
Add the words that are often mispronounced to the pronunciation dictionary at one time, which is more effective than repeatedly modifying the text.
Before cloning a voice, you must obtain written authorization from the voice rights holder, and you are also required to comply with the official usage policies.
Before launching in multiple languages, have native speakers listen to each language separately.
don't just rely on the official language list.
Latency is an official specification.
the actual value depends on the network and deployment method.
Conduct end-to-end stress testing before launch.
Refer to the official website for pricing and commercial authorization.