Cartesia's current flagship text-to-speech model; this alias tracks the newest stable Sonic 3.6 snapshot. It covers 44 languages, adding Odia and Urdu, reads Latin-script Hindi (Hinglish), and is backwards compatible with Sonic 3.5.
Pinned Sonic 3.6 snapshot released August 27, 2026 and currently the build behind the sonic-3.6 alias. Unlike the alias it never changes, which Cartesia recommends for production traffic needing consistent output.
Current Soniox text-to-speech model, generally available since August 11, 2026, adding audio tags for emotion and delivery, voice cloning and in-sentence language switching across 60+ languages; it is API-compatible with v1.
Fastest and cheapest model in Inworld's TTS-2 family, launched August 9, 2026 with a quoted 20 ms P90 time to first audio and TTS-2's language coverage and instant voice cloning, but no natural-language steering.
Audio8 TTS Preview 0.6B (July 2026) from Edge0: an open Apache 2.0 text-to-speech model with zero-shot voice cloning, a 44.1 kHz codec and a dual autoregressive design inspired by Fish Audio S2; preview release covering 11 languages.
Low-latency tier of Qwen-Audio 3.0 TTS (July 2026) with first-packet latency under 200 ms, voice cloning, and instruction and tag control of emotion, pace and style. Alibaba also lists a newer qwen-audio-3.1-tts-flash.
Alibaba's recommended speech synthesis model as of October 2026: the quality-focused Qwen-Audio 3.0 tier for audiobooks and dubbing, with voice cloning, natural-language instruction control, and cloned voices in 16 languages plus many Chinese dialects.
Murf's current streaming text-to-speech model for voice agents, launched June 2026 as the successor to Falcon. Murf quotes time to first audio under 100 ms, 150+ voices, and switching between languages within a sentence.
Microsoft AI's expressive text-to-speech model, in Azure preview since June 2026 with prebuilt voices in 15 languages and gated zero-shot voice cloning from 5-60 second clips. Superseded on 2026-10-01 by MAI-Voice-2.1 and the low-latency MAI-Voice-2.1-Flash.
Inworld's flagship text-to-speech, launched May 5, 2026, directed with bracketed natural-language steering for emotion, pacing and non-verbal sounds, plus a deliveryMode setting. Inworld quotes 100 ms P90 time to first audio and 200+ languages and locales.
Previous-generation Sonic model, superseded by Sonic 3.6 in August 2026 but still served as a stable option. Covering 42 languages, it improved on Sonic 3 in alphanumeric read-out, English heteronyms and multilingual pronunciation; the alias does not auto-upgrade to 3.6.
Only dated snapshot of Sonic 3.5, released May 4, 2026 and pinned behind the sonic-3.5 alias. It remains stable with no announced sunset; its successor is sonic-3.6-2026-08-27.
Soniox's first generally available TTS model (April 29, 2026), streaming speech in 60+ languages with accurate reading of phone numbers, emails and IDs. Deprecated by tts-rt-v2 and removed August 31, 2026, after which requests route to v2.
Preview of Soniox's first text-to-speech model, released April 23, 2026, streaming audio in 60+ languages before a sentence finishes. Since April 29, 2026 the id is an alias for tts-rt-v1, which now routes to tts-rt-v2.
Gemini 3.1 Flash TTS (gemini-3.1-flash-tts-preview), an April 2026 preview, steers delivery with inline audio tags in 70+ languages, handles two speakers and streams. Google now calls it legacy, replaced by Gemini 3.8 Flash-Lite TTS, with shutdown on November 17, 2026.
Voxtral TTS (March 2026), Mistral's first text-to-speech model: about 4B parameters, nine languages, preset voices and zero-shot cloning from a few seconds of reference audio, with roughly 70 ms model latency; open weights are CC BY-NC 4.0.
Latest alias for Mistral's Voxtral TTS, currently the 26.03 release; it speaks English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic and can clone a voice from a short sample without a transcript.
xAI's text-to-speech service at /v1/tts, generally available since March 2026, with no model selector. Covers 20 languages plus auto-detect, built-in voices (eve by default) or cloned custom voices, inline speech tags and WebSocket streaming.
Fast tier of CosyVoice 3.5 (February 2026) for custom voices only: cloned or text-designed voices with free-form instruction control, lower first-syllable latency, and real-time synthesis in 11 languages. Alibaba lists it for cost-sensitive voice design.
Current top CosyVoice model (February 2026) and Alibaba's recommended pick for voice design; synthesizes cloned or description-designed voices with free-form instruction control across 11 languages and many Chinese dialects, with no preset voice library.
