Voice AI Models in 2026: Launches by Modality
Audio is now a regular slot in the release calendar: 6 labs shipped voice or speech models since July. When everyone sells a voice, the value moves to where the calls happen.
The signal
- LaunchesMicrosoft shipped MAI-Voice-2.1 at $22 per million characters, a Flash variant at $15, and MAI-Transcribe-2-Streaming at $0.54 per audio hour (1 Oct 2026).
- LaunchesGemini 3.8 Flash TTS (23 Sept 2026): Higher-fidelity text-to-speech with voice design and 100+ languages. Audio out at $9 per million tokens until year end, $18 after (press figures).
- LaunchesNemotron 3 Diarization (23 Sept 2026): Open 100M-parameter speaker-diarization model topping VoiceArena's leaderboard.
Audio and speech made 11 of 59 named model launches since 6 Jul 2026, from 6 labs. In the lab-post sweep since 23 Sep it is 6 of 15.
Named model launches by main output, 6 Jul 2026 – 1 Oct 2026. One bar per modality.
Modality is the main output the lab announced. Launch-feed rows as listed: MAI-Voice-2.1 and its Flash variant are separate rows, so they count twice; a post naming several variants counts once. Rows before 23 Sep come from launch trackers (audio 5 of 44); rows from 23 Sep come from a sweep of lab posts that also catches smaller launches (audio 6 of 15), so the two shares are not a clean trend. The voice-out / speech-in split is our reading of each model.
Show the data (11 rows)
| Audio model | Lab | Date | Job |
|---|---|---|---|
| GPT Realtime 2.1 | OpenAI | 6 Jul 2026 | Realtime, both ways |
| Qwen-Audio-3.0-TTS | Alibaba Qwen | 20 Jul 2026 | Voice out (text to speech) |
| Grok STT 1.0 | xAI | 23 Jul 2026 | Speech in (transcribe, diarize) |
| GPT Transcribe | OpenAI | 28 Jul 2026 | Speech in (transcribe, diarize) |
| Gemini 3.8 Live | 17 Sept 2026 | Realtime, both ways | |
| Gemini 3.8 Flash TTS | 23 Sept 2026 | Voice out (text to speech) | |
| Gemini 3.8 Flash-Lite TTS | 23 Sept 2026 | Voice out (text to speech) | |
| Nemotron 3 Diarization | NVIDIA | 23 Sept 2026 | Speech in (transcribe, diarize) |
| MAI-Transcribe-2-Streaming | Microsoft | 1 Oct 2026 | Speech in (transcribe, diarize) |
| MAI-Voice-2.1 | Microsoft | 1 Oct 2026 | Voice out (text to speech) |
| MAI-Voice-2.1-Flash | Microsoft | 1 Oct 2026 | Voice out (text to speech) |
The Mode read
Microsoft launched 3 audio models today: a voice, a cheaper voice and a streaming transcriber. They are all of its launches in our feed since July. How much of the model release flow is now voice?
More than the headlines suggest. Audio and speech made 11 of 59 named launches since 6 Jul 2026, 19%, behind text and multimodal at 17 each. They came from 6 labs: OpenAI, Alibaba Qwen, xAI, Google, NVIDIA, Microsoft. 5 are voice-out models that read text aloud, 4 turn speech into text or split speakers, and 2 talk in real time both ways. In the lab-post sweep since 23 Sep, audio is 6 of 15 launches.
Voice is turning into a shelf product. 6 labs sell it, and they price it in different units: per million tokens, per million characters, per audio hour. A buyer cannot compare them at a glance, and few will try. They will buy a working phone agent from whoever already handles their calls. For voice startups, the position to hold is the call itself: the phone number, the scheduling system it books into, the compliance record it keeps. The model behind it can change every quarter.
6 labs now sell a voice. The scarce asset is the phone line it answers.
What we would back: Voice-agent companies that own a vertical’s inbound calls and the systems behind them: clinics, dealerships, home services.