Voice AI Trends in 2026: What Changed This Year
Voice AI trends in 2026: expressive models, full-duplex agents, open weights, shutdowns, funding and new labelling laws. A dated, sourced look at what changed.
The main voice AI trends of 2026 are more expressive speech models, real-time voice agents, large funding rounds, serious open-weight releases, and new rules on labelling and consent. Several figures below come from company announcements, and we label them as claims. Dates run to 17 October 2026, the date of this review.
Timeline of major 2026 voice events
| Date | Who | What | |---|---|---| | 2026-01-22 | Qwen project | Qwen3-TTS released in 0.6B and 1.7B sizes, in an Apache-2.0 open-source repository | | 2026-02-04 | ElevenLabs | $500 million Series D at an $11 billion valuation, led by Sequoia Capital | | 2026-04-15 | Google | Gemini 3.1 Flash TTS Preview launched | | 2026-06-10 | European Commission | Final voluntary Code of Practice on marking and labelling AI-generated content | | 2026-07-08 | OpenAI | GPT-Live-1 and GPT-Live-1 mini launch, able to listen and speak at the same time | | 2026-07-27 | Fish Audio | $52 million seed round announced | | 2026-08-02 | EU and California | AI Act Article 50 transparency duties and California's AI Transparency Act, as amended by AB 853, begin to apply | | 2026-08-13 | Resemble AI | Says it is no longer selling voice AI to new customers and is focusing on detection | | 2026-09-22 | Google | Gemini 3.8 Flash TTS and Flash-Lite TTS generally available in the Gemini API, with voice design and voice replication | | 2026-09-28 | ElevenLabs | Eleven v4 and Eleven v4 Turbo announced | | 2026-10-07 | Google | SynthID Detector reported as open to the public | | 2026-10-13 | Fish Audio | Drama 3 released with open weights | | 2026-11-13 | Hume AI | Scheduled end of access to its TTS and EVI APIs |
Trend 1: Expressive, directable speech
Models now take direction in plain words as well as in markup. ElevenLabs' Eleven v4, announced 2026-09-28, supports more than 90 languages and accepts inline audio tags such as "[laughs]", natural-language delivery descriptions and IPA phonemes. ElevenLabs says Eleven v4 Turbo has a median inference latency of about 100 milliseconds and about 150 milliseconds to first speech, excluding network latency. It also says the model ranked first on Artificial Analysis' Provider Voice Arena in September 2026, and that about 75% of listeners preferred it in the company's own blind tests. We did not check the leaderboard page, so treat these as vendor claims.
Google's Gemini 3.8 Flash TTS adds voice design from a text description and says it supports over 100 languages. Cartesia's Sonic 3.6 snapshot, marked stable on 2026-08-27, supports 44 languages, including Odia and Urdu.
OndaLabs works in the same direction. You control delivery by describing it in plain words and with supported expressive tags in the script, and Onda Flash is the most expressive model, available on paid plans. You can try it on the text to speech page.
Trend 2: Real-time and full-duplex voice agents
OpenAI launched GPT-Live-1 and GPT-Live-1 mini on 2026-07-08 as models that listen and speak at the same time, according to Reuters. GPT-Live-1 can hand reasoning and tool use to a backend agent. OpenAI bills it at $0.05 per minute, billed per second, with backend usage billed separately.
ElevenLabs says it handled over 15 million weekly agent conversations by 2026-09-30, tripled since February. That is a company-reported figure. Voice agents are clearly growing, but the sources here do not show that voice is replacing chat.
Trend 3: Money and consolidation
- ElevenLabs funding. The company announced a $500 million Series D at an $11 billion valuation on 2026-02-04. It says it has raised $781 million across five rounds since 2022.
- ElevenLabs revenue. In May 2026 the company said it passed $500 million in annual recurring revenue within the first four months of 2026. This is company-reported.
- ElevenLabs tender offer. On 2026-09-30 it announced a $300 million employee tender offer at a $22 billion valuation. The figure comes from the company, and we found no independent source for it.
- Fish Audio. It announced a $52 million seed round on 2026-07-27, with company-reported figures of about $21 million in ARR and over 8 million users.
- Meta and PlayAI. Meta acquired the voice startup PlayAI in July 2025, according to TechCrunch. Terms were not disclosed.
- Hume AI. Hume is sunsetting its TTS and EVI APIs. Access ends on 2026-11-13 at 12:01 a.m. EST, and account data is permanently deleted after that date. Its notice does not mention the Octave model by name. See the Hume AI Octave review.
- Resemble AI. The company says it is not selling voice AI to new customers. It continues to support existing customers and now sells detection, watermarking, identity and intelligence products.
- PlayHT. Some competitors report that PlayHT has closed, but we found no official notice. See PlayHT alternatives before you choose a provider.
Trend 4: Open weights get serious
- Qwen3-TTS (released 2026-01-22) comes in 0.6B and 1.7B sizes, with VoiceDesign, CustomVoice (nine preset speakers) and Base (voice cloning) variants. It covers 10 languages, and its repository is Apache-2.0. The README cites a "3-second rapid voice clone" and latency as low as 97 milliseconds. Both are vendor claims.
- Fish Audio Drama 3 (released 2026-10-13) has open weights on Hugging Face and covers 83 languages. Its Community License allows free commercial use under US$500,000 in annual revenue. Above that, or to offer it as a TTS, cloning or inference service, you need a licence. The licence prohibits cloning a real person's voice without consent.
- Chatterbox from Resemble AI is MIT-licensed, and every output carries the Perth watermark. Resemble says it keeps Chatterbox and DramaBox open under "Voice AI Research".
Open weights are cheap to try, but each comes with its own licence terms, and those terms decide what you can build.
Trend 5: Provenance and consent get built in
Google says every clip from its Gemini audio models carries SynthID, and that replicated voices carry C2PA credentials. Voice replication through Gemini 3.8 TTS needs a 30-second sample plus a verbal consent recording from the voice owner. Google says the feature is not available through AI Studio in Illinois, Texas, the EEA, the UK, Switzerland or India.
The C2PA specification reached version 2.4 in April 2026. ElevenLabs offers an Audio Detector that checks its own watermarks first. For how these marks work and where they fail, read AI audio watermarking.
Trend 6: New rules on voice
| Rule | Status as of October 2026 | Key date | |---|---|---| | EU AI Act Article 50 | Applies | From 2026-08-02. Law-firm summaries say generative systems already on the market have until 2026-12-02 for Article 50(2) marking. | | California AI Transparency Act (AB 853) | Applies to covered providers | Operative 2026-08-02. Platforms from 2027-01-01. Capture devices from 2028-01-01. | | Washington SSB 5886 | In effect | Effective 2026-06-11 | | NO FAKES Act of 2026 | Bill | Reintroduced 2026-05-20 | | Denmark deepfake bill | Bill | Due to be introduced 2026-10-08, with a target start in 2027 |
The NO FAKES Act of 2026 (S.4591 in the Senate, H.R.8915 in the House) went through a Senate Judiciary Committee markup in June 2026 and is still a bill as of October 2026.
This is general information, not legal advice.
What to do about it
- Creators. Check the commercial terms on your plan, disclose AI narration where platforms ask, and use recordings you have permission to use.
- Developers. Compare licences before you build on an open model. Apache-2.0, MIT and community licences set different limits. Plan for shutdowns, as the Hume date shows, and export your data before any deadline.
- Teams. Do not use a voice alone to approve payments or grant access. Sam Altman's 2025 remark that AI "has fully defeated" voiceprint authentication was an opinion, but it points to the same risk.
- Everyone. Keep records of how each voice was made and what permission covers it.
Where OndaLabs fits
OndaLabs is the new name of NeoSpeech, and it is an AI voice studio. It offers text to speech, two-speaker dialogue, speech to text, custom voices, image generation and music generation. It has 30 prebuilt voices that work in every supported language, and 37 language pages, such as /text-to-speech/spanish. Plans run from a free plan at $0 to Business at $199 a month. Onda Lite is available on all plans, and Onda Flash is available on paid plans. You can compare them on the pricing page.
Frequently asked questions
What are the biggest voice AI trends in 2026?
Three stand out: expressive models that take direction in plain words, real-time voice agents such as OpenAI's GPT-Live-1, and new rules on labelling and consent. Many figures come from the vendors themselves, so check each one before you rely on it.
Which new text to speech models launched in 2026?
Examples include Qwen3-TTS in January, Google's Gemini 3.8 Flash TTS and Flash-Lite TTS in September, ElevenLabs' Eleven v4 in September and Fish Audio Drama 3 in October. Each has its own languages and licence terms, so check the current pages before you choose one.
Are voice agents replacing chat?
The sources in this roundup do not show that. ElevenLabs reports over 15 million weekly agent conversations, which is a company-reported figure. No source compares voice agent use with chat use.
What happened to PlayHT and Hume?
Hume AI has scheduled the end of its TTS and EVI API access for 2026-11-13, and account data is deleted after that date. For PlayHT, some competitors report a closure, but we found no official notice, so verify it before you rely on it.
What new laws cover AI voices in 2026?
Washington's forged digital likeness law took effect on 2026-06-11. EU Article 50 and California's AI Transparency Act began to apply on 2026-08-02. The NO FAKES Act of 2026 and Denmark's deepfake bill are still bills.