# Voix à voix : les meilleurs agents vocaux

QUELLE IA, https://quelleia.com/voix/voix-a-voix/, relevé du 9 novembre 2026 (édition 2026-W46). Classements à part du score des modèles de langage : chaque tableau vient d'une seule source, dans son unité, sans score combiné. Méthode : https://quelleia.com/demarche/#voix.

## Indice voix à voix d'Artificial Analysis

GPT-Live 1 (OpenAI) arrive premier avec un indice de 84,3 sur 100, sur 12 modèles (Artificial Analysis, relevé du 9 novembre 2026). Mesure : indice Speech to Speech d'Artificial Analysis, sur 100 : moyenne à poids égaux de leurs mesures de raisonnement à l'écoute (Big Bench Audio), de travail d'agent au téléphone (τ-Voice), de préférence des votants dans leur arène d'agents vocaux et de taux de réussite des tâches, selon leur méthode ; seuls les modèles mesurés sur toutes ces épreuves y figurent. Indice sur 100 : plus il est haut, mieux c'est. Rang partagé à valeur égale au dixième, comme la source l'affiche ; elle ne publie pas de marge.

Source : Artificial Analysis, https://artificialanalysis.ai/speech-to-speech (relevé du 9 novembre 2026).

| Rang | Modèle | Éditeur | Indice | Nom chez la source |
| --- | --- | --- | --- | --- |
| 1 | [GPT-Live 1](https://quelleia.com/voix/modeles/openai-gpt-live-1/) | OpenAI | 84,3 | GPT-Live-1 (Astra, medium) |
| 2 | [Qwen Audio 3.0 Realtime Plus](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-plus/) | Alibaba | 84,1 | Qwen Audio 3.0 Realtime Plus |
| 3 | [Grok Voice 2.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-2-0/) | SpaceXSI | 82,9 | Grok Voice 2.0 |
| 4 | [GPT-Realtime-2.1](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1/) | OpenAI | 79,1 | GPT-Realtime-2.1 High |
| 5 | [GPT-Realtime-2](https://quelleia.com/voix/modeles/openai-gpt-realtime-2/) | OpenAI | 77,2 | GPT-Realtime-2 (High) |
| 6 | [Qwen Audio 3.0 Realtime Flash](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-flash/) | Alibaba | 76,3 | Qwen Audio 3.0 Realtime Flash |
| 7 | [Grok Voice 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-1-0/) | SpaceXSI | 75,7 | Grok Voice 1.0 |
| 8 | [GPT-Realtime-1.5](https://quelleia.com/voix/modeles/openai-gpt-realtime-1-5/) | OpenAI | 72,0 | GPT-Realtime-1.5 |
| 9 | [Gemini 3.1 Flash Live Preview](https://quelleia.com/voix/modeles/google-gemini-3-1-flash-live-preview/) | Google | 69,5 | Gemini 3.1 Flash Live Preview - High |
| 10 | [GPT-Realtime](https://quelleia.com/voix/modeles/openai-gpt-realtime/) | OpenAI | 69,2 | GPT Realtime (Aug '25) |
| 11 | [GPT-Realtime-2.1 Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1-mini/) | OpenAI | 59,0 | GPT-Realtime-2.1 Mini Minimal |
| 12 | [GPT-Realtime Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-mini/) | OpenAI | 58,1 | GPT Realtime Mini (Oct '25) |

## Raisonnement à l'écoute, Big Bench Audio

StepAudio 3 Realtime (StepFun) arrive premier avec 99,7 % de réussite, sur 38 modèles (Artificial Analysis, relevé du 9 novembre 2026). Mesure : part de bonnes réponses, dites à voix haute, à 1 000 questions de raisonnement posées à l'oral en anglais (Big Bench Audio, tiré de Big Bench Hard : sophismes, orientation, dénombrement, vrai ou faux), notées par un modèle juge. Part de réussite : plus elle est haute, mieux c'est. Rang partagé à valeur égale au dixième, comme la source l'affiche ; elle ne publie pas de marge.

Source : Artificial Analysis, https://artificialanalysis.ai/speech-to-speech (relevé du 9 novembre 2026).

| Rang | Modèle | Éditeur | Réussite | Nom chez la source |
| --- | --- | --- | --- | --- |
| 1 | [StepAudio 3 Realtime](https://quelleia.com/voix/modeles/stepaudio-3-realtime/) | StepFun | 99,7 % | StepAudio 3 Realtime |
| 2 | [Qwen Audio 3.0 Realtime Plus](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-plus/) | Alibaba | 99,2 % | Qwen Audio 3.0 Realtime Plus |
| 3 | [Phonic 1 (High)](https://quelleia.com/voix/modeles/phonic-1-high/) | Phonic | 98,9 % | Phonic 1 (High) |
| 4 | [Eleven Realtime v1 Max](https://quelleia.com/voix/modeles/elevenlabs-eleven-realtime-v1-max/) | ElevenLabs | 98,8 % | Eleven Realtime v1 Max |
| 5 | [Qwen3.5-Omni-Plus-Realtime](https://quelleia.com/voix/modeles/alibaba-qwen3-5-omni-plus-realtime/) | Alibaba | 98,7 % | Qwen3.5 Omni Plus Realtime |
| 6 | [Grok Voice 3.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-3-0/) | SpaceXSI | 98,5 % | Grok Voice 3.0 |
| 7 | [Gemini 3.8 Live Extended Thinking](https://quelleia.com/voix/modeles/google-gemini-3-8-live-extended-thinking/) | Google | 97,7 % | Gemini 3.8 Live Extended Thinking (High) |
| 8 | [Step-Audio R1.1 (Realtime)](https://quelleia.com/voix/modeles/step-audio-r1-1-realtime/) | StepFun | 97,6 % | Step-Audio R1.1 (Realtime) |
| 9 | [Grok Voice 2.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-2-0/) | SpaceXSI | 97,2 % | Grok Voice 2.0 |
| 10 | [Grok Voice 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-1-0/) | SpaceXSI | 97,1 % | Grok Voice 1.0 |
| 11 | [Gemini 3.1 Flash Live Preview](https://quelleia.com/voix/modeles/google-gemini-3-1-flash-live-preview/) | Google | 96,6 % | Gemini 3.1 Flash Live Preview - High |
| 11 | [GPT-Realtime-2](https://quelleia.com/voix/modeles/openai-gpt-realtime-2/) | OpenAI | 96,6 % | GPT-Realtime-2 (High) |
| 13 | [Qwen Audio 3.0 Realtime Flash](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-flash/) | Alibaba | 96,1 % | Qwen Audio 3.0 Realtime Flash |
| 14 | [GPT-Realtime-2.1](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1/) | OpenAI | 95,8 % | GPT-Realtime-2.1 High |
| 15 | [Grok Voice Fast 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-fast-1-0/) | SpaceXSI | 93,3 % | Grok Voice Fast 1.0 |
| 16 | [Gemini 3.8 Live](https://quelleia.com/voix/modeles/google-gemini-3-8-live/) | Google | 91,7 % | Gemini 3.8 Live |
| 17 | [Gemini 2.5 Flash Native Audio Dialog Thinking](https://quelleia.com/voix/modeles/google-gemini-2-5-flash-native-audio-dialog-thinking/) | Google | 90,7 % | Gemini 2.5 Flash Native Audio Dialog Thinking |
| 18 | [GPT-Live 1](https://quelleia.com/voix/modeles/openai-gpt-live-1/) | OpenAI | 90,1 % | GPT-Live-1 (Astra, medium) |
| 19 | [Amazon Nova 2 Sonic](https://quelleia.com/voix/modeles/amazon-nova-2-sonic/) | Amazon | 88,1 % | Nova 2.0 Sonic (Mar 2026) |
| 20 | [Deepslate Opal](https://quelleia.com/voix/modeles/deepslate-opal/) | Deepslate | 85,0 % | Deepslate Opal |
| 21 | [GPT-Realtime](https://quelleia.com/voix/modeles/openai-gpt-realtime/) | OpenAI | 83,3 % | GPT Realtime (Aug '25) |
| 22 | [GPT-Realtime-1.5](https://quelleia.com/voix/modeles/openai-gpt-realtime-1-5/) | OpenAI | 81,4 % | GPT-Realtime-1.5 |
| 23 | [Eleven Realtime v1 Mini](https://quelleia.com/voix/modeles/elevenlabs-eleven-realtime-v1-mini/) | ElevenLabs | 75,9 % | Eleven Realtime v1 Mini |
| 24 | [GPT-Realtime-2.1 Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1-mini/) | OpenAI | 74,7 % | GPT-Realtime-2.1 Mini High |
| 25 | [GPT-4o Mini Realtime](https://quelleia.com/voix/modeles/openai-gpt-4o-mini-realtime/) | OpenAI | 68,9 % | GPT-4o mini Realtime (Dec '24) |
| 26 | [Gemini 2.5 Flash Native Audio Dialog](https://quelleia.com/voix/modeles/google-gemini-2-5-flash-native-audio-dialog/) | Google | 68,6 % | Gemini 2.5 Flash Native Audio Dialog |
| 26 | [Higgs Realtime](https://quelleia.com/voix/modeles/boson-ai-higgs-realtime/) | Boson AI | 68,6 % | Higgs Realtime |
| 28 | [GPT-Realtime Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-mini/) | OpenAI | 63,6 % | GPT Realtime Mini (Oct '25) |
| 29 | [Qwen3.5-Omni-Flash-Realtime](https://quelleia.com/voix/modeles/alibaba-qwen3-5-omni-flash-realtime/) | Alibaba | 59,0 % | Qwen3.5 Omni Flash Realtime |
| 30 | [Qwen3 Omni 30B A3B Instruct (flash)](https://quelleia.com/voix/modeles/alibaba-qwen3-omni-30b-a3b-instruct-flash/) | Alibaba | 58,7 % | Qwen3 Omni 30B A3B Instruct (flash) |
| 31 | [Raon-SpeechChat-9B](https://quelleia.com/voix/modeles/krafton-raon-speechchat-9b/) | KRAFTON | 57,5 % | Raon SpeechChat |
| 32 | [Qwen3 Omni 30B A3B Instruct (realtime)](https://quelleia.com/voix/modeles/alibaba-qwen3-omni-30b-a3b-instruct-realtime/) | Alibaba | 56,8 % | Qwen3 Omni 30B A3B Instruct (realtime) |
| 33 | [GPT-4o Audio](https://quelleia.com/voix/modeles/openai-gpt-4o-audio/) | OpenAI | 54,3 % | GPT-4o audio chatcompletions |
| 34 | [Freeze-Omni](https://quelleia.com/voix/modeles/vita-freeze-omni/) | VITA | 33,4 % | Freeze-Omni |
| 35 | [Nemotron Voicechat](https://quelleia.com/voix/modeles/nvidia-nemotron-voicechat/) | NVIDIA | 27,0 % | Nemotron Voicechat |
| 36 | [PersonaPlex](https://quelleia.com/voix/modeles/nvidia-personaplex-7b-v1/) | NVIDIA | 19,1 % | Personaplex 7B v1 |
| 37 | [FLM-Audio](https://quelleia.com/voix/modeles/cofe-ai-flm-audio/) | Cofe AI | 16,0 % | FLM-Audio |
| 38 | [Moshi](https://quelleia.com/voix/modeles/kyutai-moshi/) | Kyutai | 4,4 % | Moshi |

## Tenue de la conversation, Full Duplex Bench

StepAudio 3 Realtime (StepFun) arrive premier avec 98,9 % de réussite, sur 31 modèles (Artificial Analysis, relevé du 9 novembre 2026). Mesure : part des situations de conversation bien tenues, en anglais : ne pas couper la parole pendant une pause, prendre la parole au bon moment, répondre à une interruption, continuer malgré un « mm-hmm » (sous-ensemble de Full Duplex Bench v1 et v1.5 choisi par Artificial Analysis). Part de réussite : plus elle est haute, mieux c'est. Rang partagé à valeur égale au dixième, comme la source l'affiche ; elle ne publie pas de marge.

Source : Artificial Analysis, https://artificialanalysis.ai/speech-to-speech (relevé du 9 novembre 2026).

| Rang | Modèle | Éditeur | Réussite | Nom chez la source |
| --- | --- | --- | --- | --- |
| 1 | [StepAudio 3 Realtime](https://quelleia.com/voix/modeles/stepaudio-3-realtime/) | StepFun | 98,9 % | StepAudio 3 Realtime |
| 2 | [Qwen Audio 3.0 Realtime Plus](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-plus/) | Alibaba | 98,4 % | Qwen Audio 3.0 Realtime Plus |
| 3 | [GPT-Live 1](https://quelleia.com/voix/modeles/openai-gpt-live-1/) | OpenAI | 97,3 % | GPT-Live-1 (Sol, low) |
| 4 | [Grok Voice 3.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-3-0/) | SpaceXSI | 97,0 % | Grok Voice 3.0 |
| 4 | [Qwen Audio 3.0 Realtime Flash](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-flash/) | Alibaba | 97,0 % | Qwen Audio 3.0 Realtime Flash |
| 6 | [Gemini 3.8 Live](https://quelleia.com/voix/modeles/google-gemini-3-8-live/) | Google | 96,1 % | Gemini 3.8 Live |
| 6 | [GPT-Realtime-2](https://quelleia.com/voix/modeles/openai-gpt-realtime-2/) | OpenAI | 96,1 % | GPT-Realtime-2 (Minimal) |
| 8 | [Eleven Realtime v1 Mini](https://quelleia.com/voix/modeles/elevenlabs-eleven-realtime-v1-mini/) | ElevenLabs | 95,8 % | Eleven Realtime v1 Mini |
| 9 | [GPT-Realtime Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-mini/) | OpenAI | 95,7 % | GPT Realtime Mini (Oct '25) |
| 9 | [GPT-Realtime-1.5](https://quelleia.com/voix/modeles/openai-gpt-realtime-1-5/) | OpenAI | 95,7 % | GPT-Realtime-1.5 |
| 9 | [GPT-Realtime-2.1](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1/) | OpenAI | 95,7 % | GPT-Realtime-2.1 High |
| 12 | [Grok Voice 2.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-2-0/) | SpaceXSI | 95,1 % | Grok Voice 2.0 |
| 13 | [Eleven Realtime v1 Max](https://quelleia.com/voix/modeles/elevenlabs-eleven-realtime-v1-max/) | ElevenLabs | 94,1 % | Eleven Realtime v1 Max |
| 14 | [GPT-Realtime](https://quelleia.com/voix/modeles/openai-gpt-realtime/) | OpenAI | 93,9 % | GPT Realtime (Aug '25) |
| 15 | [Higgs Realtime](https://quelleia.com/voix/modeles/boson-ai-higgs-realtime/) | Boson AI | 92,8 % | Higgs Realtime |
| 16 | [Gemini 3.8 Live Extended Thinking](https://quelleia.com/voix/modeles/google-gemini-3-8-live-extended-thinking/) | Google | 91,9 % | Gemini 3.8 Live Extended Thinking (High) |
| 17 | [GPT-Realtime-2.1 Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1-mini/) | OpenAI | 91,8 % | GPT-Realtime-2.1 Mini Minimal |
| 18 | [PersonaPlex](https://quelleia.com/voix/modeles/nvidia-personaplex-7b-v1/) | NVIDIA | 91,0 % | Personaplex 7B v1 |
| 19 | [GPT-4o Realtime](https://quelleia.com/voix/modeles/openai-gpt-4o-realtime/) | OpenAI | 89,8 % | GPT-4o Realtime (Dec 2024) |
| 20 | [Raon-SpeechChat-9B](https://quelleia.com/voix/modeles/krafton-raon-speechchat-9b/) | KRAFTON | 86,2 % | Raon SpeechChat |
| 21 | [Deepslate Opal](https://quelleia.com/voix/modeles/deepslate-opal/) | Deepslate | 85,7 % | Deepslate Opal |
| 22 | [Grok Voice 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-1-0/) | SpaceXSI | 77,8 % | Grok Voice 1.0 |
| 23 | [Gemini 3.1 Flash Live Preview](https://quelleia.com/voix/modeles/google-gemini-3-1-flash-live-preview/) | Google | 74,3 % | Gemini 3.1 Flash Live Preview - High |
| 24 | [Qwen3 Omni 30B A3B Instruct (flash)](https://quelleia.com/voix/modeles/alibaba-qwen3-omni-30b-a3b-instruct-flash/) | Alibaba | 72,7 % | Qwen3 Omni 30B A3B Instruct (flash) |
| 25 | [Grok Voice Fast 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-fast-1-0/) | SpaceXSI | 71,6 % | Grok Voice Fast 1.0 |
| 26 | [FLM-Audio](https://quelleia.com/voix/modeles/cofe-ai-flm-audio/) | Cofe AI | 62,0 % | FLM-Audio |
| 27 | [Moshi](https://quelleia.com/voix/modeles/kyutai-moshi/) | Kyutai | 61,0 % | Moshi |
| 28 | [Freeze-Omni](https://quelleia.com/voix/modeles/vita-freeze-omni/) | VITA | 58,7 % | Freeze-Omni |
| 29 | [Nemotron Voicechat](https://quelleia.com/voix/modeles/nvidia-nemotron-voicechat/) | NVIDIA | 52,9 % | Nemotron Voicechat |
| 30 | [Gemini 2.5 Flash Native Audio Preview (Dec 2025)](https://quelleia.com/voix/modeles/google-gemini-2-5-flash-native-audio-preview-12-2025/) | Google | 44,0 % | Gemini 2.5 Flash Native Audio Preview (Dec 2025) |
| 31 | [Gemini 2.5 Flash Native Audio Preview (Sep 2025)](https://quelleia.com/voix/modeles/google-gemini-2-5-flash-native-audio-preview-09-2025/) | Google | 30,3 % | Gemini 2.5 Flash Native Audio Preview (Sep 2025) |

## Délai avant le premier son

Raon-SpeechChat-9B (KRAFTON) arrive premier avec un délai de 0,04 s avant le premier son, sur 34 modèles (Artificial Analysis, relevé du 9 novembre 2026). Mesure : temps moyen, en secondes, entre la fin de la question et le premier son de la réponse, sur les questions de Big Bench Audio. Délai en secondes : plus il est court, plus l'agent répond vite. Rang partagé à valeur égale au centième de seconde.

Source : Artificial Analysis, https://artificialanalysis.ai/speech-to-speech (relevé du 9 novembre 2026).

| Rang | Modèle | Éditeur | Délai | Nom chez la source |
| --- | --- | --- | --- | --- |
| 1 | [Raon-SpeechChat-9B](https://quelleia.com/voix/modeles/krafton-raon-speechchat-9b/) | KRAFTON | 0,04 s | Raon SpeechChat |
| 2 | [Deepslate Opal](https://quelleia.com/voix/modeles/deepslate-opal/) | Deepslate | 0,44 s | Deepslate Opal |
| 3 | [Eleven Realtime v1 Max](https://quelleia.com/voix/modeles/elevenlabs-eleven-realtime-v1-max/) | ElevenLabs | 0,57 s | Eleven Realtime v1 Max |
| 4 | [Phonic 1 (High)](https://quelleia.com/voix/modeles/phonic-1-high/) | Phonic | 0,61 s | Phonic 1 (High) |
| 5 | [Eleven Realtime v1 Mini](https://quelleia.com/voix/modeles/elevenlabs-eleven-realtime-v1-mini/) | ElevenLabs | 0,62 s | Eleven Realtime v1 Mini |
| 6 | [Gemini 2.5 Flash Native Audio Dialog](https://quelleia.com/voix/modeles/google-gemini-2-5-flash-native-audio-dialog/) | Google | 0,64 s | Gemini 2.5 Flash Native Audio Dialog |
| 7 | [Grok Voice 2.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-2-0/) | SpaceXSI | 0,70 s | Grok Voice 2.0 |
| 8 | [Grok Voice Fast 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-fast-1-0/) | SpaceXSI | 0,78 s | Grok Voice Fast 1.0 |
| 9 | [Qwen3.5-Omni-Flash-Realtime](https://quelleia.com/voix/modeles/alibaba-qwen3-5-omni-flash-realtime/) | Alibaba | 0,79 s | Qwen3.5 Omni Flash Realtime |
| 10 | [GPT-Realtime Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-mini/) | OpenAI | 0,81 s | GPT Realtime Mini (Oct '25) |
| 10 | [GPT-Realtime-1.5](https://quelleia.com/voix/modeles/openai-gpt-realtime-1-5/) | OpenAI | 0,81 s | GPT-Realtime-1.5 |
| 10 | [Grok Voice 3.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-3-0/) | SpaceXSI | 0,81 s | Grok Voice 3.0 |
| 13 | [GPT-Realtime-2.1 Mini](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1-mini/) | OpenAI | 0,85 s | GPT-Realtime-2.1 Mini Minimal |
| 14 | [Qwen3 Omni 30B A3B Instruct (realtime)](https://quelleia.com/voix/modeles/alibaba-qwen3-omni-30b-a3b-instruct-realtime/) | Alibaba | 0,88 s | Qwen3 Omni 30B A3B Instruct (realtime) |
| 15 | [Gemini 3.1 Flash Live Preview](https://quelleia.com/voix/modeles/google-gemini-3-1-flash-live-preview/) | Google | 0,96 s | Gemini 3.1 Flash Live Preview - Minimal |
| 16 | [GPT-Realtime-2.1](https://quelleia.com/voix/modeles/openai-gpt-realtime-2-1/) | OpenAI | 0,97 s | GPT-Realtime-2.1 Minimal |
| 17 | [GPT-Realtime](https://quelleia.com/voix/modeles/openai-gpt-realtime/) | OpenAI | 0,98 s | GPT Realtime (Aug '25) |
| 18 | [GPT-Realtime-2](https://quelleia.com/voix/modeles/openai-gpt-realtime-2/) | OpenAI | 1,12 s | GPT-Realtime-2 (Minimal) |
| 19 | [Amazon Nova 2 Sonic](https://quelleia.com/voix/modeles/amazon-nova-2-sonic/) | Amazon | 1,14 s | Nova 2.0 Sonic (Mar 2026) |
| 20 | [Gemini 3.8 Live](https://quelleia.com/voix/modeles/google-gemini-3-8-live/) | Google | 1,18 s | Gemini 3.8 Live |
| 21 | [GPT-Live 1](https://quelleia.com/voix/modeles/openai-gpt-live-1/) | OpenAI | 1,24 s | GPT-Live-1 (Sol, low) |
| 22 | [Grok Voice 1.0](https://quelleia.com/voix/modeles/spacexai-grok-voice-1-0/) | SpaceXSI | 1,25 s | Grok Voice 1.0 |
| 23 | [GPT-4o Mini Realtime](https://quelleia.com/voix/modeles/openai-gpt-4o-mini-realtime/) | OpenAI | 1,27 s | GPT-4o mini Realtime (Dec '24) |
| 24 | [Gemini 3.8 Live Extended Thinking](https://quelleia.com/voix/modeles/google-gemini-3-8-live-extended-thinking/) | Google | 1,35 s | Gemini 3.8 Live Extended Thinking (High) |
| 25 | [Higgs Realtime](https://quelleia.com/voix/modeles/boson-ai-higgs-realtime/) | Boson AI | 1,47 s | Higgs Realtime |
| 26 | [GPT-4o Realtime](https://quelleia.com/voix/modeles/openai-gpt-4o-realtime/) | OpenAI | 1,49 s | GPT-4o Realtime (Dec 2024) |
| 27 | [Step-Audio R1.1 (Realtime)](https://quelleia.com/voix/modeles/step-audio-r1-1-realtime/) | StepFun | 1,53 s | Step-Audio R1.1 (Realtime) |
| 28 | [Qwen Audio 3.0 Realtime Plus](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-plus/) | Alibaba | 1,54 s | Qwen Audio 3.0 Realtime Plus |
| 29 | [Qwen Audio 3.0 Realtime Flash](https://quelleia.com/voix/modeles/alibaba-qwen-audio-3-0-realtime-flash/) | Alibaba | 1,55 s | Qwen Audio 3.0 Realtime Flash |
| 30 | [Qwen3.5-Omni-Plus-Realtime](https://quelleia.com/voix/modeles/alibaba-qwen3-5-omni-plus-realtime/) | Alibaba | 2,64 s | Qwen3.5 Omni Plus Realtime |
| 31 | [GPT-4o Audio](https://quelleia.com/voix/modeles/openai-gpt-4o-audio/) | OpenAI | 3,38 s | GPT-4o audio chatcompletions |
| 32 | [Gemini 2.5 Flash Native Audio Dialog Thinking](https://quelleia.com/voix/modeles/google-gemini-2-5-flash-native-audio-dialog-thinking/) | Google | 3,87 s | Gemini 2.5 Flash Native Audio Dialog Thinking |
| 33 | [Qwen3 Omni 30B A3B Instruct (flash)](https://quelleia.com/voix/modeles/alibaba-qwen3-omni-30b-a3b-instruct-flash/) | Alibaba | 4,82 s | Qwen3 Omni 30B A3B Instruct (flash) |
| 34 | [StepAudio 3 Realtime](https://quelleia.com/voix/modeles/stepaudio-3-realtime/) | StepFun | 8,84 s | StepAudio 3 Realtime |

## Annoncés, pas encore mesurés

Aucune source publique ne les classe encore : ils n'entrent dans aucun tableau. Leur fiche reprend ce que l'éditeur publie.

- [Azure Realtime](https://quelleia.com/voix/modeles/microsoft-azure-realtime/) (Microsoft), voix à voix, annoncé le 25 juillet 2026
- [Fun-Audio-Chat-8B](https://quelleia.com/voix/modeles/fun-audio-chat-8b/) (FunAudioLLM), voix à voix, annoncé le 23 décembre 2025
- [Hydra](https://quelleia.com/voix/modeles/smallest-ai-hydra/) (Smallest AI), voix à voix, annoncé le 23 mai 2026
- [LFM3-Audio-2B](https://quelleia.com/voix/modeles/lfm3-audio-2b/) (Liquid AI), voix à voix, annoncé le 27 octobre 2026
- [MiniCPM-o 4.5](https://quelleia.com/voix/modeles/minicpm-o-4-5/) (OpenBMB), voix à voix, annoncé le 17 mai 2026