OpenAI2024
GPT-4o Audio
Spoken dialoguecommercialstreamemotionvendor demo
End-to-end spoken dialogue, not a vocoder bolted onto a chatbot.
09 · 2024–
LLM / spoken dialogue
The model is not reading a script. It is talking, with tools, interruptions, and a voice that is part of the language model.
Playable examples
Sky’s story clip from OpenAI’s 2023 announcement. Hume’s sarcasm read is Octave performing a script. Sesame’s conversation clip is CSM, not a vocoder reciting.
GPT-4o Audio (2024) and the later Realtime line, including GPT-Realtime-2 in 2026, generate speech as a native modality. You can still force them to recite, but that is not what they are for. Sky, in the 2023 announcement clips, is already closer to a performer than to a vocoder.
Moshi (Kyutai) is full-duplex: it listens and speaks on overlapping streams. Sesame’s CSM is a conversational speech model with open 1B weights. Dia generates multi-speaker scripts in one pass. ChatTTS aims at fillers and chat cadence. Hume Octave reads for meaning and shifts tone without tags. Amazon Nova Sonic and Gemini Live sit on the same side of the line.
Gemini 3.1 Flash TTS (April 2026) is a useful counterexample: a Gemini-native recitation model with 200+ audio tags and no streaming. It is an LLM that still knows it is reading. The live-agent path at Google is a different API.
On this plate
Sky’s story clip from OpenAI’s 2023 announcement. Hume’s sarcasm read is Octave performing a script. Sesame’s conversation clip is CSM, not a vocoder reciting.
Sesame2025
Sesame CSM
Spoken dialogueopenclonestreamemotionvendor demo
The ‘this doesn’t sound like TTS’ conversational demo.
Hume AI2024
Hume Octave 2
Spoken dialoguecommercialclonestreamemotionvendor demo
Emotion as a first-class modelling target, not a slider.
Microsoft2025
VibeVoice
Spoken dialogueresearchcloneemotionvendor demo
Context length as the product.
Kyutai2024
Kyutai Moshi
Spoken dialogueopenstreamemotionno clip
Overlapping conversation, not turn-taking TTS.
Google DeepMind2026
Gemini 3.1 Flash TTS
Spoken dialoguecommercialemotionno clip
Leaderboard-era Gemini speech with SynthID watermarking.