Voice AI: How AI voices moved from the lab into every hotline

From robot voices to ElevenLabs: how the breakthrough happened

Computers that read text aloud have been around for a long time. Classic speech synthesis (text-to-speech, or TTS) stitched voices together from recorded snippets. The result was understandable but monotonous, like sat-nav devices or station announcements. In 2016 Google's DeepMind lab showed with WaveNet that a neural network can generate the sound wave directly, and speech sounded much more natural. In 2018 Google demonstrated with Duplex how an AI could call a hair salon by itself. Yet good artificial voices long remained in the hands of a few corporations such as Google, Amazon and Microsoft.

That changed with ElevenLabs. It was founded in 2022 by Piotr Dąbkowski, a former Google engineer, and Mati Staniszewski, formerly at the data company Palantir. Both are from Poland and cite the often poorly dubbed films of their childhood as inspiration. In January 2023 a first public test version went online: type in text, pick a voice, and out came a surprisingly lively audio file. With a few minutes of your own recordings you could even recreate your own voice (voice cloning).

The business figures mostly come from the company itself and have not been independently verified. According to ElevenLabs, its valuation was 9 million US dollars at its first funding in 2022 and 11 billion in February 2026. In September 2026 investors bought employee shares at a price valuing the company at 22 billion dollars. In September 2026, chief executive Mati Staniszewski put annualised recurring revenue (ARR) at 600 million dollars, up from more than 500 million in May. According to ElevenLabs, about 55 percent of revenue comes from businesses. Deutsche Telekom is among its customers and investors.

Context: valuations in employee share sales are not stock market prices. They show what individual investors are willing to pay and can change quickly.

ElevenLabs' new models and how they compare

The current flagship is Eleven v4, available since 28 September 2026, together with the faster variant Eleven v4 Turbo. As with its predecessor, emotion and pacing can be steered with instructions in square brackets such as [whispers] or [laughs], and dialogue with several speakers is possible. According to ElevenLabs both speak more than 90 languages, including German. A voice recorded in one language can then speak fluently in another. The Turbo variant is said to respond in about 100 milliseconds and is meant for phone agents. According to the maker, 10 seconds of audio are now enough for a quick voice clone.

Other building blocks include Flash v2.5, a cheap and very fast reading model, the speech recognition model Scribe v2 for more than 90 languages, with its own version for doctor-patient conversations, and the agent platform ElevenAgents for complete phone assistants. Since August 2025 Eleven Music has also generated songs, now in version 2.5, according to the company trained on licensed production music, with further licensing deals with the music publisher Kobalt and with Merlin, which represents independent record labels.

How good is it really? The best-known independent measurement is the Speech Arena run by the analysis firm Artificial Analysis. People listen to two voices reading the same text without knowing which model is behind them and pick the more natural one. Thousands of such duels produce an Elo rating, as in chess. As of 10 October 2026, Eleven v4 Turbo leads with 1327 points and Eleven v4 with 1323. Close behind are Alibaba's Qwen-Audio-3.1 (1298), Cartesia Sonic 3.6 (1284) and Google's Gemini 3.8 Flash (1272). OpenAI's best pure reading model on the list dates from 2023 and trails far behind at 1105 points. Hume, known for emotive voices, scores 1051.

Context: the lead is narrow and changes quickly. Compared with its predecessor Eleven v3 (1176 points), v4 is a big step up. English voices were tested; there is no comparably large leaderboard for German. ElevenLabs' claim that around 75 percent of listeners prefer v4 over rivals comes from its own tests.

From reading aloud to making calls: voice agent platforms

A nice voice alone does not make a conversation. That takes three building blocks: speech recognition that turns your words into text, a language model like the one behind ChatGPT that formulates a reply, and a voice that speaks it. From 2023 onwards, companies emerged that assemble exactly this chain for phone calls. Experts speak of voice agents.

The best-known example is Vapi from San Francisco, a start-up from the Y Combinator accelerator that launched its platform publicly in 2024. According to Reuters, Vapi was valued at 130 million dollars in December 2024. In May 2026 Vapi said it raised 50 million dollars in fresh funding, at a valuation of around 500 million according to a source cited by TechCrunch. By its own account Vapi has handled more than one billion calls. Amazon's doorbell brand Ring reportedly routes all incoming customer calls through Vapi.

Similar platforms include Retell AI, Bland and Synthflow. There are also specialists for individual building blocks: Deepgram for speech recognition, Cartesia for especially fast voices, Hume for voices with emotional expression and Sesame for the most natural conversational voices possible. At the same time the large AI companies built speech directly into their models. OpenAI launched the Realtime API in October 2024, letting developers build conversations without going through text, and Google offers something comparable with Gemini Live. How you can talk to these assistants yourself is explained in a separate article on voice mode.

Falling prices and open models: free, but with a catch

As of October 2026, voice AI is no longer a niche. Experts speak of commoditisation: a product becomes an interchangeable mass good, and competition runs on price. When OpenAI introduced the Realtime API in October 2024, one minute of AI-spoken output cost about 24 US cents, according to the company. By our conversion of the official prices, it is now around 8 cents on the large model and just over 2 cents on the smaller one.

Even cheaper are open models (open source or open weights): anyone can download the model files and run them on their own computer. Of 96 models on the Artificial Analysis leaderboard, 16 are open. Some examples:

Kokoro is tiny at 82 million parameters, runs even on ordinary laptops and is under the Apache licence, so it may also be used commercially. It cannot clone voices, though, and scores 1064 points on the leaderboard. Chatterbox from Resemble AI is under the permissive MIT licence, speaks 23 languages including German and clones a voice from about 10 seconds of audio. Every recording gets an inaudible digital watermark. Fish Audio S2 Pro (1116 points) and the top-ranked open model Breeze TTS 2 (1222 points, English and Chinese only), by contrast, are free to use only for research and non-commercial purposes. Voxtral TTS from the French company Mistral, which speaks German, is likewise released for non-commercial use only.

The quality: the best open model sits about 100 Elo points below the top, and the commercially usable ones considerably further. In blind tests the leaders clearly win more often, but for reading apps or hobby projects open models are often good enough.

The catch is misuse. Anyone running an open model on their own computer bypasses a provider's safeguards and can clone any voice from a short video. Watermarks like Chatterbox's only help as long as nobody removes them from the code. Important for you: in Germany, recreating and spreading another person's voice without their consent generally violates their personality rights, even if the tool is freely available.

What this means for you in everyday life

You notice the change most clearly on the phone. More and more hotlines at insurers, telecom firms or delivery services use AI agents that speak freely, understand questions and handle simple requests themselves. In Germany there are also AI phone assistants for doctors' practices that book appointments or take prescription requests. The upside: less time on hold. The downside: with unusual requests the AI may misunderstand you. In that case, explicitly ask for a human.

What the technology can do for people who are ill is especially moving. US Congresswoman Jennifer Wexton lost her clear speech to the neurological disease PSP. Her team sent ElevenLabs more than an hour of old speeches, and within a few days a clone of her former voice was ready. On 25 July 2024, according to NPR, she used it to give the first speech in the US House of Representatives with an AI-cloned voice. ElevenLabs says it has pledged to provide such voices free of charge to up to one million affected people, for example with ALS.

Films and games are changing too. In May 2025 Darth Vader returned to the game Fortnite with the voice of James Earl Jones, who died in 2024, generated by AI with help from ElevenLabs and with his family's consent. Players could chat with him freely. Shortly after launch a streamer got the character to swear, and Epic Games had to fix it.

In Germany a dispute escalated in early 2026. Netflix wanted contracts stating that recordings by German dubbing actors could be used to train AI. According to the German public broadcaster's news site Tagesschau, around 80 percent of dubbing actors had stopped working for Netflix by February 2026, and well-known German voices were missing from Stranger Things. The German voice actors' association VDS demands that speakers be allowed to refuse and be paid for AI training. In early September 2026 it asked Germany's Federal Cartel Office to examine whether Netflix is abusing its market power, and filed a data protection complaint in Berlin. There is no decision yet; speakers who do not sign still get no Netflix work. AI voices are also gaining ground in audiobooks; when buying, check whether a human or an AI is narrating.

For accessibility the development is often a gain: natural-sounding reading voices help people with visual impairments or reading difficulties.

The dark side: voice clones, fraud and labelling rules

The same technology that makes audiobooks cheaper also makes fraud easier. If 10 seconds of audio are enough to imitate a voice, a video from social media is often all it takes. Fraudsters can sound like a family member who urgently needs money. One well-known case came from the US: in January 2024 voters received automated calls with a fake voice of President Joe Biden. According to audio experts it was created with ElevenLabs.

Official figures now show the scale of the problem for the first time. In its internet crime report for 2025, the US Federal Bureau of Investigation (FBI) includes a dedicated section on AI for the first time: 22,364 complaints with an AI link and losses of almost 893 million dollars. In so-called distress scams alone, where fraudsters pose as relatives in trouble, victims reported losses of more than 5 million dollars with a likely AI link. The real numbers are probably higher because many cases go unreported.

Many people in Germany underestimate the risk. In the 2026 Cybersicherheitsmonitor of Germany's Federal Office for Information Security (BSI), only just over half of respondents (54 percent) thought it technically possible for criminals to use AI to imitate a relative's voice on the phone. A simple defence: if someone asks for money, hang up and call the person back on the number you know. Some families also agree on a code word. How to spot fakes is explained in a separate article.

Legally, Article 50 of the EU AI Act has applied since 2 August 2026. It requires AI systems that talk directly to people to be designed so that you learn you are talking to an AI, unless this is obvious. Artificially generated voices that seem real (so-called deepfakes) must be disclosed as such. For machine-readable marking of AI audio, systems already on the market before then have a transition period until 2 December 2026. Fraudsters, however, do not follow such rules. Labelling mainly helps with reputable providers.