ElevenLabs

ElevenLabs builds AI audio models and products for generating speech, cloning voices, translating spoken performances and running conversational agents. Founded in 2022 by Mati Staniszewski and Piotr Dąbkowski, it organizes its products into ElevenAgents, ElevenCreative and ElevenAPI: an agent platform, a media creation environment and developer access to audio models. As of September 2026, Staniszewski is CEO and Dąbkowski is CTO.[1][6]

Enterprise agents became a major investment priority in 2026. With its February Series D, ElevenLabs announced improvements to turn-taking and expressive conversational synthesis, alongside further research into emotional conversational models, dubbing and audio general intelligence. This broader platform grew out of earlier work on speech synthesis and multilingual dubbing.[6][16]

The founders were high-school friends in Poland; Dąbkowski later worked at Google and Staniszewski at Palantir. In a September 2024 interview, Staniszewski recalled working with Dąbkowski at Opera and continuing private weekend hackathons after their careers diverged. A 2021 voice-analysis experiment suggested improvements to a speaker’s delivery from a recording. Months later, Dąbkowski proposed returning to audio after watching a film with Poland’s single-reader lektor narration. Their first pitch deck focused exclusively on dubbing.[2][1]

Staniszewski recalled early investors questioning the founders’ audio experience and the size of the dubbing market before Credo Ventures and angels provided backing. He attributed the subsequent expansion beyond dubbing to choosing markets where the technology was sufficiently developed and customers already needed it.[2]

ElevenLabs launched its beta platform in January 2023. Its June Series A announcement introduced Studio for creating and editing long-form spoken content, using speech synthesis and audio infilling, and an AI Speech Classifier for identifying ElevenLabs-generated audio. Studio’s editing workflow included speaker assignment, pause insertion and segment regeneration.[3][16]

On October 10, 2023, ElevenLabs launched AI Dubbing, translating audio and video into more than 20 languages while aiming to retain the original speaker’s voice, emotion and intonation. The tool combined multilingual synthesis, voice cloning, text and audio processing, speaker detection and separation of dialogue from background music and noise.[16]

ElevenLabs also supplied synthesis to applications built by other companies. In a September 2023 Wondercraft interview, Youssef Rizk confirmed a workflow that turned generated or supplied scripts into speech using ElevenLabs. Wondercraft provided pronunciation controls, podcast hosting and production defaults, and was developing templates; its professional-translator quality assurance belonged to Wondercraft’s dubbing workflow.[22]

The January 2024 product announcements added editing and creator compensation around synthesis. Forthcoming Dubbing Studio controls covered transcripts, translations and timecodes. A Voice Library marketplace would pay creators when others used their verified professional voice replicas, with creators controlling availability and compensation terms; a small alpha group was already earning. By January 2025, ElevenLabs reported more than 5,000 shared voices and more than $2 million in cumulative creator rewards.[4][5]

Eleven v3 debuted as a research preview on June 3, 2025. It introduced multi-speaker dialogue and inline audio tags such as [excited], [whispers] and [sighs]. The preview required more careful prompting and carried reliability and latency limitations; ElevenLabs recommended Turbo or Flash for realtime conversation at that stage. A dated changelog announced v3 API access on August 20, 2025, followed by general availability on February 2, 2026.[11][12][13]

The general-availability release addressed interpretation as well as delivery: phone numbers could become large quantities, chemical notation such as SO₂ could be misread, and sports scores could sound like subtraction. ElevenLabs reported errors falling from 15.3% to 4.9% on an internal interpretation benchmark covering 27 categories across eight languages. That measurement concerns the company’s numbers-and-symbols test, rather than general synthesis accuracy or a comparison with competing models.[13]

The model lineup serves different application requirements. Eleven v3 supports expressive, multi-speaker speech across more than 70 languages; Multilingual v2 is positioned for stable long-form output across 29 languages; Flash v2.5 targets fast interactive synthesis across 32 languages. ElevenLabs quotes approximately 75 milliseconds of model latency for Flash v2.5, excluding application and network latency. This component estimate does not describe a caller’s complete wait for an agent response.[8]

Flash v2.5 disables default number normalization to preserve low latency. Telephone numbers, dates and currencies may therefore be spoken unexpectedly unless the application converts them into suitable spoken text. ElevenLabs recommends preprocessing with a language model, choosing Multilingual v2 when normalization matters, or enabling the supported enterprise normalization parameter.[8]

The official Python SDK exposes these choices directly: synthesis requests supply text, a voice identifier, a model identifier and an output format. Streaming returns audio chunks as generation proceeds for incremental playback or processing, while instant cloning accepts sample recordings. The client library is MIT-licensed.[10]ⁿ¹

ElevenAgents coordinates four components: speech recognition, a selected or custom language model, speech synthesis and a proprietary turn-taking model for conversation timing. Workflows, tool calls, interruption and timeout controls, testing, analytics, and telephony, web and mobile integrations surround that pipeline. Recognizing a request, deciding how to respond, speaking and managing when to yield are separate responsibilities.[9]

The speech-recognition lineup includes Scribe v2, with word timestamps, speaker diarization, dynamic audio tagging, entity detection and keyterm prompting, and Scribe v2 Realtime for streaming partial transcripts. The SDK’s Speech Engine illustrates the developer-controlled side of the pipeline: a server receives transcripts and streams language-model responses back for synthesis. When the user interrupts, the previous asynchronous response handler is automatically cancelled.[8][10]

Eleven Music launched on August 5, 2025, generating vocal or instrumental music from natural-language prompts with controls over genre, style and structure, plus editing of individual sections or whole songs. Music API access for paid users followed on August 20. Music v2, announced May 26, 2026, improved selective inpainting—regenerating a chosen part of a track—and expanded section-by-section composition.[14][12][15]

The original AI Dubbing product aimed to preserve the speaker’s voice, emotion and intonation. Dubbing v2, announced May 28, 2026, conditions directly on the original performance to preserve tone, pacing, delivery and emotional intent. Its sync-aware translation adjusts phrasing and timing around the original starts and stops. For professional productions, ElevenProductions combines the model with human translators, voice casting and audio mixing.[17][16]

The Impact Program, announced August 20, 2024, offered free cloning and text-to-speech access to people with ALS/MND through Bridging Voice and The Scott-Morgan Foundation. Recordings supply a digital voice replica usable in assistive communication software and devices, while the partners provide onboarding and support. Bridging Voice confirms free ElevenLabs cloning for anyone with ALS in the United States. In January 2025, ElevenLabs reported more than 1,000 participants with speech impairments; its ambition to empower one million voices remained a goal.[18][19][5]

The program announcement describes former NFL player Tim Green using old recordings, with Bridging Voice’s help, to create a replica for his podcast Nothing Left Unsaid after losing his ability to speak. The announcement also discloses podcast sponsorship negotiations with ElevenLabs.[18]

Voice replication also created impersonation risks. Associated Press reported early beta-period misuse and ElevenLabs’ decision to replace anonymous free cloning access with a payment-information requirement.[21]

ElevenLabs’ safety policy describes blocking celebrity and other high-risk voice cloning, technological verification for Professional Voice Cloning, prerelease red-teaming, monitoring with classifiers and human reviewers, and enforcement against prohibited uses. Its AI Speech Classifier targets ElevenLabs-generated audio. The company acknowledges that safeguards can block legitimate users or miss malicious ones; the professional-cloning verification requirement is specific to that tool.[20]

ElevenLabs announced a $19 million Series A on June 20, 2023, co-led by Nat Friedman, Daniel Gross and Andreessen Horowitz, and an $80 million Series B on January 22, 2024, co-led by the same investors. The company reported more than one million registered users in June 2023 and growth from five to 40 employees by January 2024. Registrations measure neither active use nor paying customers.[3][4]

On January 30, 2025, ElevenLabs announced a $180 million Series C co-led by a16z and ICONIQ Growth at a $3.3 billion valuation, bringing total funding to $281 million. It reported more than 250,000 agents built and use of its platform or API by employees at over 60% of Fortune 500 companies. Those figures describe agent creation and employee use, rather than production deployments or enterprise customer penetration. Its research expansion included a Warsaw R&D center and work on Indic language coverage in India.[5]

On February 4, 2026, ElevenLabs announced $500 million raised in a Series D led by Sequoia at an $11 billion valuation, with Andrew Reed joining the board. Total funding reached $781 million across five rounds. Staniszewski expressed an intention to build toward an IPO, without announcing a schedule.[6]

A May 5, 2026 founder update, subsequently updated August 31, reported surpassing $500 million ARR during the first four months of 2026. It named additional investors at the third close of the Series D, including BlackRock, Wellington and NVIDIA through NVentures, and reported a completed $100 million tender offer and 530 teammates across more than 50 countries. The tender offer is distinct from primary fundraising; ARR is a company-reported annualized metric, rather than audited annual revenue. The staffing figure belongs to that dated update.[7]