Cartesia2024
Cartesia Sonic 3.5
Streamingcommercialclonestreamemotionvendor demo
The lowest published time-to-first-audio among flagship APIs.
10 · 2023–
Streaming / low-latency
Time-to-first-audio is the number voice agents actually feel. Architectures that decode left-to-right, or that scale linearly, win here.
Playable examples
You cannot hear latency on a file. You can hear the models people chose for streaming: Pocket TTS and Piper on the bench line, CosyVoice 2 as a clone, Aura-2 and Sonic as labelled vendor exhibits.
A streaming TTS system emits audible audio before the full sentence is known. That requires incremental text, a model that does not need the future, and a transport (often WebSocket) that does not wait for a complete MP3.
Cartesia Sonic uses a state-space model so cost scales linearly with length; Sonic 3.5’s published figures sit around 40–82 ms to first audio. Inworld Mini, Deepgram Aura-2, ElevenLabs Flash v2.5, and CosyVoice 2 occupy the same brief. Piper is the offline version of the idea: small enough that ‘streaming’ is just ‘faster than real time on a Pi.’
Quality and latency still trade. ElevenLabs v3 is the expressive model; Flash is the one you put on a phone call. Google is explicit that Gemini 3.1 Flash TTS does not stream. If your product is an agent, that sentence is the whole review.
On this plate
You cannot hear latency on a file. You can hear the models people chose for streaming: Pocket TTS and Piper on the bench line, CosyVoice 2 as a clone, Aura-2 and Sonic as labelled vendor exhibits.
Deepgram2024
Deepgram Aura-2
Streamingcommercialstreamvendor demo
The other half of a Deepgram voice stack.
Kyutai2026
Kyutai Pocket TTS
Streamingopenclonestreambench clip
Open voice cloning that actually fits in a laptop CPU budget.
FunAudioLLM / Alibaba2024
CosyVoice 2
Streamingopenclonestreamemotionvendor demo
Open streaming clone quality that is actually usable live.
Open Home Foundation2023
Piper
End-to-end VAEopenstreambench clip
Edge and offline deployment; hundreds of community voices.
ElevenLabs2022
ElevenLabs
Codec LMcommercialclonestreamemotionvendor demo
Voice library, cloning, and inline performance tags like [whispers].