AI voice tools used to mean text-to-speech that read a script aloud. In 2026 the headline shift has been towards real-time voice: models that listen, reason and speak in one conversation, built mainly for phone agents and live assistants. Rules on cloned voices and synthetic audio have also moved.
This article lists the dated developments we found from January to October 2026, separates vendor claims from independent evidence, and ends with questions to ask before you put a voice tool in front of customers. Some items come from newsletters and recaps rather than vendor pages, and we say so.
Key takeaways
- OpenAI launched GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper in its API in early May 2026, and a recap lists GPT-Live-1 full-duplex models on 8 July.
- ElevenLabs published Eleven Realtime on 20 October 2026: two speech-to-speech models for voice agents, with benchmarks measured by Artificial Analysis and no pricing on the page we read.
- Funding was heavy: a recap reports Deepgram raised $130 million on 13 January and ElevenLabs a $500 million Series D on 4 February.
- EU transparency rules for synthetic audio were reported to apply from 2 August 2026, and a revised NO FAKES Act was reported as reintroduced in May.
- Most performance figures are published by the vendors themselves, so run your own call tests in your own accents and languages.
Summary
Voice AI in 2026 split into two stories. The product story is faster, more conversational speech models that can use tools mid-call. The business story is money and regulation: large funding rounds, and new or advancing rules on synthetic audio and likeness. For a small business the realistic uses are transcription, voice notes, phone answering and translation, each with its own risks.
If you are new to the category, read our beginner’s guide to AI voice and text-to-speech first.
Timeline
| Date (2026) | Event | Source type |
|---|---|---|
| 13 January | Deepgram reported to have raised $130 million at a $1.3 billion valuation | Recap article |
| 4 February | ElevenLabs reported to have raised a $500 million Series D at an $11 billion valuation | Recap article |
| 15 April | Google released Gemini 3.1 Flash TTS | Recap article |
| 7 or 8 May | OpenAI launched GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper | News article (8 May) and recap (7 May) |
| May | Revised NO FAKES Act reported as reintroduced | Recap article |
| 8 July | OpenAI launched GPT-Live-1 full-duplex models | Recap article |
| 2 August | EU transparency rules for synthetic audio reported to take effect | Recap article |
| 20 October | ElevenLabs published Eleven Realtime | Vendor blog |
| 19 to 24 October | A week of launches from Salesforce, Zoom, IBM, AWS and others | Newsletter |
The shift to real-time voice
Older voice assistants chained three steps: speech to text, a language model, then text to speech. Each step adds delay. The 2026 launches try to cut that delay and make conversations feel natural, including handling interruptions. A recap describes OpenAI’s GPT-Live-1 models as full-duplex, meaning they listen and speak at the same time. A newsletter lists Block’s Dessa as another full-duplex model, in research preview on 23 October.
ElevenLabs takes a different design. Its blog says Eleven Realtime uses a cascaded architecture with components tuned end to end, with each step producing text. The company says this supports visibility and lets its platform apply guardrails and personal data redaction. Different designs trade speed, transparency and control differently, so there is no single right answer.
OpenAI voice models
The Next Web reported on 8 May 2026 that OpenAI released three voice models in its API. Another recap gives 7 May, so confirm the date on OpenAI’s own page.
| Model | What the report says | Reported price |
|---|---|---|
| GPT-Realtime-2 | Audio in and out in one model, with reasoning, parallel tool calls, spoken preambles while tools run, a 128K context window and a reasoning effort setting | $32 per million audio input tokens and $64 per million audio output tokens |
| GPT-Realtime-Translate | Live translation, more than 70 input and 13 output languages | $0.034 per minute |
| GPT-Realtime-Whisper | Streaming speech-to-text for low-latency transcription | $0.017 per minute |
These prices are from a news report, not an official page we opened. Checked on 11 November 2026 only as reported, so confirm on the vendor page before you budget. OpenAI’s own benchmark claims, such as gains on Big Bench Audio, and customer results like Zillow’s were quoted in the report as company figures. We could not verify them. The report also says OpenAI offers EU data residency.
ElevenLabs Eleven Realtime
The ElevenLabs blog, published 20 October 2026 and updated 29 October, describes Eleven Realtime as a new class of speech-to-speech models for voice agents.
- Two models: v1 Mini for low-cost natural conversations, and v1 Max for complex multi-step tasks and heavy tool calling.
- More than 90 languages and over 17,000 voices from the ElevenLabs library.
- One model ID for audio in and audio out, compatible with OpenAI’s Realtime API, with WebSocket connection only.
- Used best through ElevenAgents, which the page says adds phone, web, WhatsApp and contact-centre channels, workflows, a knowledge base, guardrails, testing and monitoring.
- The page does not state pricing.
The page cites Artificial Analysis benchmarks: for Max, 98.8 percent on speech reasoning (Big Bench Audio) and 0.57 seconds to first audio, and for Mini 75.9 percent and 0.62 seconds. These were presented by the vendor on its own blog, so treat them as vendor-presented. The page also says the standalone Realtime API lacks the platform features that ElevenAgents provides.
Other October launches
The voiceaispace.com newsletter for 19 to 25 October 2026 lists many launches and links to what it says are original sources. We did not open those sources, so treat the list as reported.
| Date | Company | Item as reported |
|---|---|---|
| 19 October | Salesforce AI Research | Launa-Omni, a real-time enterprise voice model that shows each reply’s text before it is spoken |
| 19 October | monday.com | Harmony, voice agents for sales calls |
| 20 October | Mistral AI | Voxtral 3 speech models and a voice agent SDK |
| 20 October | Zoom | Voice Agent, text-to-speech and Interpreter APIs opened to developers |
| 21 October | IBM | Granite Speech 5.0, led by an open 2B transcription and translation model under Apache 2.0 |
| 22 October | Amazon Web Services | Amazon Polly generative engine upgrade with 14 new voices |
| 23 October | Deepgram | Nova-3 multilingual expanded to 55 languages in one model |
Rules and consent
A voxbooster mid-year recap lists several regulatory items. Treat the list as a pointer and check primary sources. It says that:
- A revised NO FAKES Act was reintroduced in May 2026 to create a federal right against unauthorised AI voice and likeness replicas.
- The TAKE IT DOWN Act compliance deadline of 19 May required platforms to remove flagged content within 48 hours.
- EU AI Act Article 50 transparency rules for synthetic audio take effect on 2 August, with fines up to 15 million euros or 3 percent of global turnover.
- Denmark advanced a bill to protect a person’s face and voice for 50 years after death.
Our article on new AI regulation covers the EU dates from law firm sources. For the basics on cloned voices, read voice cloning consent and legal basics. This is general information, not legal advice.
What it means for tool buyers
Pick the use first
Transcription of your own meetings, voice notes turned into drafts, phone answering and live translation each have different needs. A transcription tool that is accurate on your accent matters more for notes than for a sales call. See how to combine voice notes and AI for writing for a low-risk use.
Check recording and privacy settings
Voice data is personal data. A newsletter item says Twilio made account-wide default call recording settings generally available in one region, a reminder that recording is a setting you control. Check consent rules for recording calls where you and your callers are located. How to protect company data when using AI tools applies to audio as much as text.
Be honest with callers
Transparency rules are tightening, and plain disclosure is also good practice. Tell callers when they are speaking to an AI.
How to test a voice tool
- Record twenty real calls or clips with your accents, background noise and names.
- Measure transcription errors on names, numbers and product terms, not on easy sentences.
- Test interruptions and silence.
- Test what happens when the system cannot help: does it hand over to a person?
- Check where audio and transcripts are stored and for how long.
- Compare cost per minute at your real volume.
The steps match our guide on testing an AI tool before you pay for it.
What to watch next
- Independent comparisons of real-time voice models in languages beyond English.
- Whether the NO FAKES Act advances in Congress.
- How EU transparency rules are applied to synthetic audio in practice.
- Published pricing for newer models such as Eleven Realtime.
Frequently asked questions
What is a speech-to-speech model?
A model that takes audio in and produces audio out directly or through tightly linked steps, to cut the delay of chaining separate speech-to-text, language and text-to-speech systems.
What does full-duplex mean?
The system can listen and speak at the same time, which helps with interruptions. A recap uses the term for OpenAI’s GPT-Live-1.
Are the vendor benchmarks reliable?
They are claims published by the vendors. Some cite a third party, such as Artificial Analysis, but you should still test on your own audio.
Do I need consent to clone a voice?
Generally you should get explicit permission. Rules vary. See our guide on voice cloning consent and take legal advice for commercial use.
What does a voice agent cost?
It depends on the model and platform. OpenAI’s reported prices are per million tokens or per minute. ElevenLabs did not state pricing on the page we read.
Conclusion
The big change this year is conversational, tool-using voice. The sensible response is to pick a narrow use, test with your own audio and check consent and disclosure. Begin with the beginner’s guide, then review voice cloning consent if cloned voices are in scope.
Sources
- The Next Web, OpenAI GPT-Realtime-2 voice models, 8 May 2026, accessed 11 November 2026
- ElevenLabs, Eleven Realtime, 20 October 2026, accessed 11 November 2026
- Voxbooster, Voice AI 2026 mid-year recap, accessed 11 November 2026
- VoiceAISpace, Voice AI news, 19 to 25 October 2026, accessed 11 November 2026
