AI ConsultingAboutVideosBookFAQContact
DE|EN
Book Free Consultation

The short version

On October 9, 2026, HeyGen launched its own voice model, HeyGen Voice, and its press release says it debuts at #1 on Artificial Analysis' leaderboard. That holds for one of two lists: on the Controlled Voice Arena, where every model clones the same eight English voices, HeyGen Voice was at rank 1 to 2 on October 11, shared with Qwen. On the main leaderboard, Eleven v4 Turbo and Eleven v4 are on top, HeyGen Voice isn't listed there, and it has no German rating.

So I cloned my own voice in HeyGen Voice and in ElevenLabs v4, both as instant clones from the same 2:13-minute recording, with default settings and the first take of every sentence. Then came four English rounds and a German bonus round as a blind test, each time against my real voice.

By my ears, HeyGen was closer to my voice in two English rounds and ElevenLabs in the other two, so 2:2. The German bonus round went to ElevenLabs, which makes it 3:2. That's very close, and these were first takes only: with two, three or four tries per sentence you'd probably get a perfect version from either tool. I'd be happy with both.

What else stood out: ElevenLabs steers the delivery with audio tags like [laughs] in the text, and I'd wish for that in HeyGen as well. HeyGen, on the other hand, includes AI avatars in the plan, a huge bonus if you work with avatar videos.

What you'll take away

  • What HeyGen's “#1” actually refers to, and why Artificial Analysis runs two leaderboards
  • How a fair comparison is set up: same recording, default settings, first take, loudness-matched blind test
  • Which clone came closer to my real voice in each round, in English and in German
  • What audio tags like [laughs] do in ElevenLabs, and how the same sentence sounds in HeyGen without tags
  • What both tools cost per month and which plan includes the voice clone
  • A tip for HeyGen if generating fails without a clear reason

Is HeyGen Voice #1? The two leaderboards

HeyGen's press release of October 9, 2026 is titled “HeyGen Launches HeyGen Voice, Debuting at #1 on Artificial Analysis' Leaderboard”. It doesn't say which leaderboard. Artificial Analysis compares speech models in blind tests: raters hear two clips of the same text in random order, pick the better one, and only see the model names afterwards. Loudness is normalized, and the rankings are recalculated several times a day. There are two separate leaderboards, and they measure different things:

  • Provider Voice Arena, the main list under “Text to Speech Leaderboard”: every model speaks with its provider's own stock voices. This measures catalog voices.
  • Controlled Voice Arena: every model clones the same eight professionally recorded voices (4 US, 4 UK) and reads the same text. This measures voice cloning. This is the list where HeyGen Voice is on top.
ModelVoice cloning, EnglishMain leaderboardVoice cloning, German
HeyGen Voice (HeyGen)1201 points, rank 1 to 2, 1,531 samplesnot listednot listed
Qwen-Audio-3.1-TTS-Plus (Alibaba)1187, rank 1 to 21298, rank 3 to 4not in the top 13
Eleven v4 Turbo (ElevenLabs)1168, rank 3 to 41327, rank 1 to 21416, rank 1 to 2
Eleven v4 (ElevenLabs)1162, rank 3 to 41323, rank 1 to 21414, rank 1 to 2

Source: Artificial Analysis, Controlled Voice Leaderboard (English and German) and Provider Voice Leaderboard, retrieved October 11, 2026, the day of the recording. The scores change all the time.

How to read the table

  • Scores only count within one leaderboard. Each list is calculated separately and has its own reference point. That's why Eleven v4 Turbo has 1168 points on one list and 1327 on the other.
  • Rank 1 to 2 means a shared top spot. Next to the position, Artificial Analysis shows a rank range and an uncertainty for each score (±16 points for HeyGen Voice). HeyGen Voice and Qwen both have the range 1 to 2, so they share the top spot.
  • HeyGen Voice had the fewest samples of the top six: 1,531, compared with 1,921 to 2,778 for the other five. It has only been listed since October, so its score can still move.
  • German: on the German list of the Controlled Voice Arena, Eleven v4 Turbo and Eleven v4 lead. HeyGen Voice has no rating there.

How I tested

The test compares what the Controlled Voice Arena measures: voice clone against voice clone. Both are instant clones.

  • One recording for both. 2:13 minutes of free speech, recorded with a Shure SM7B through a Scarlett 2i2. HeyGen got the WAV file, ElevenLabs a 320 kbps MP3 of the same file, because ElevenLabs accepts at most 10 MB per file for an instant clone. On length: HeyGen's API docs call 1 to 3 minutes ideal; ElevenLabs recommends about 1 to 2 minutes and advises against more than 3.
  • The right engine. In HeyGen, the clone ran on HeyGen Voice, as shown in the voice list. HeyGen also offers ElevenLabs as an engine; that would have turned the test into ElevenLabs against ElevenLabs. In ElevenLabs, the clone ran on Eleven v4.
  • Default settings. No sliders changed, no audio tags in the blind rounds. Background noise removal was on by default in both. In ElevenLabs, default means stability 50, similarity 75 (v4 has no speed slider).
  • First take counts. HeyGen offered three variants of the clone, and I picked one. ElevenLabs gives two takes per sentence, and I always took the first.
  • Blind test. My real voice, HeyGen and ElevenLabs in random order, numbered and matched for loudness so the louder file doesn't win. Claude Code built and shuffled the blind test from my files, so I didn't know the order myself until the reveal.
  • How the samples were made. In ElevenLabs I generated the sentences directly with the clone. In HeyGen I made them as scenes of an avatar video; on the Creator plan I could download the audio on its own. Tip: use a full-frame avatar look for this. With a circle look, generating failed twice for me, and the error message didn't say why.
  • Also measured: the clone was ready after about 22 seconds in HeyGen and about 11 seconds in ElevenLabs.

The test sentences, identical for both tools

  • Clips from the start (series sentence): “I built this entire project over a single weekend and honestly I still cannot believe it actually works.” I've cloned the same sentence before, in Voicebox vs ElevenLabs and in ElevenLabs Part 1, with different settings there.
  • Blind test 1, neutral: “Hey, it's Lukas. Today I'm testing whether a ten-second voice clone can really replace two hours of studio recordings. I'll keep it honest: if it works, I'll say so, and if it doesn't, you'll hear why.”
  • Blind test 2, name and price: “ElevenLabs charges $22 a month, Claude Code costs more, and my name is Lukas Hüttis.”
  • Blind test 3, tutorial tone: “Open the settings, click on Voices and paste your script. The whole process takes about two minutes, and you can export the result as an MP3 file.”
  • German bonus round: “Hey, hier ist Lukas. Heute teste ich, ob ein Stimmklon aus zehn Sekunden wirklich zwei Stunden Studioaufnahmen ersetzen kann. Ich bleibe ehrlich: Wenn es funktioniert, sage ich es, und wenn nicht, hörst du, warum.”
  • Expression, not blind: ElevenLabs with audio tags: “[laughs] Okay, I did not expect that. [whispers] Don't tell anyone, but this might change how I record. [sighs] Or I'll be re-recording everything tomorrow.” HeyGen got the same sentence without tags.

Results round by round

Judged by my own ears: which clone is closer to my real voice? First takes only, and overall it was very close.

RoundOrder in the videoCloser to my voice
Clips from the start1 HeyGen, 2 real, 3 ElevenLabsHeyGen
Blind test 1: neutral1 ElevenLabs, 2 real, 3 HeyGenElevenLabs
Blind test 2: name and price1 real, 2 HeyGen, 3 ElevenLabsHeyGen
Blind test 3: tutorial1 HeyGen, 2 ElevenLabs, 3 realElevenLabs
German bonus round1 HeyGen, 2 ElevenLabs, 3 realElevenLabs
Final scoreEnglish 2:2, with the German round 3:2ElevenLabs, very close

Expression with audio tags (not blind): ElevenLabs reads instructions like [laughs] or [whispers] in the text and acts them out; HeyGen doesn't support tags like these, so it read the sentence without them. HeyGen's version already sounded pretty close to my voice. I love how ElevenLabs handles the tags, and I'd wish for that in HeyGen as well.

Prices compared

As of October 11, 2026, billed monthly.

HeyGenElevenLabs
Voice clone in a planCreator, $29/month, voice cloning includedInstant clone from Starter, $6/month
Professional clone$99/month add-on according to the press release; not on the pricing page on October 10from Creator, $22/month
Free plan1 voice clone (according to the help center)no instant cloning
Also in the testAI avatars and avatar videos includedAudio tags like [laughs] for delivery
API per 1M characters$30 (HeyGen Voice)$80 (Eleven v4), $40 (Eleven v4 Turbo)

Billed annually, HeyGen Creator costs $24/month according to the pricing page (retrieved October 10, 2026). API prices according to Artificial Analysis, retrieved October 11, 2026. When I recorded the video, Eleven v4 was on a free two-week promotion. For current prices, check the HeyGen and ElevenLabs pricing pages.

Sources

Everything on this page with its retrieval date. Provider details come from official pages, help centers and documentation.

Artificial Analysis

HeyGen

ElevenLabs

Frequently asked questions

HeyGen Voice is HeyGen's in-house voice model, launched on October 9, 2026. According to the API docs, the instant clone is made from a single recording, of which only the first 3 minutes are used; on October 10 it was still listed there as a preview. For the professional clone, the docs require at least 20 minutes of audio. According to the docs, HeyGen Voice clones and speaks 39 languages. Audio tags for laughing or whispering, like those in ElevenLabs v4, are not documented. HeyGen also offers other engines, including ElevenLabs. In my test, the voice list showed “HeyGen Voice” for my clone, and it was ready after about 22 seconds.

On one of Artificial Analysis' two leaderboards. On the Controlled Voice Arena, where every model clones the same eight English voices, HeyGen Voice led with 1201 points on October 11, 2026. Artificial Analysis shows a rank range of 1 to 2 for it, shared with Qwen-Audio-3.1 (1187 points), and HeyGen Voice had the fewest samples of the top six. On the main leaderboard, the Provider Voice Arena, where each model uses its provider's own voices, HeyGen Voice isn't listed; Eleven v4 Turbo (1327) and Eleven v4 (1323) lead there. HeyGen Voice has no German rating. In my own blind test it was 2:2 after the four English rounds and 3:2 for ElevenLabs after the German bonus round, judged by my ears and with first takes only.

An instant clone needs very little, but both providers recommend more than the minimum. HeyGen's upload dialog asked for at least 10 seconds on October 11, its help center says at least 30 seconds, and its API docs call 1 to 3 minutes ideal and only use the first 3 minutes. ElevenLabs' blog mentions 10 seconds, but its docs recommend about 1 to 2 minutes and advise against more than 3. For a professional clone, HeyGen's API docs require at least 20 minutes (the press release says 30 minutes to 3 hours), ElevenLabs at least 30 minutes, ideally 2 to 3 hours. In this test, I created both instant clones from the same 2:13-minute recording.

According to HeyGen's API docs, yes: HeyGen Voice clones and speaks 39 languages, including German, Spanish, French and Japanese. There's no independent score for German yet, though: on Artificial Analysis, HeyGen Voice has no German rating. On the German list of the Controlled Voice Arena, Eleven v4 Turbo (1416 points) and Eleven v4 (1414) led on October 11, 2026. ElevenLabs lists more than 90 languages for the v4 family. In my German bonus round, ElevenLabs was closer to my voice by my ears, but not by as much as I'd expected.

As of October 11, 2026, billed monthly: HeyGen's Creator plan costs $29/month with voice cloning included, with 40 voice clones according to the help center; billed annually, the pricing page lists $24/month. The free plan includes one voice clone. HeyGen's press release lists the professional clone as a $99/month add-on. ElevenLabs offers instant cloning from Starter ($6/month) and professional cloning from Creator ($22/month). For API use, Artificial Analysis lists HeyGen Voice at $30 per 1 million characters, Eleven v4 at $80 and Eleven v4 Turbo at $40. HeyGen's plans also include AI avatars.

It can. In HeyGen you choose the voice engine: HeyGen's help center lists “ElevenLabs V3, Orca, Panda, Fish” in that menu (article last updated September 24, 2026), and HeyGen's API docs name the engine “orca” HeyGen Voice and offer ElevenLabs models, including Eleven v4, through the same speech endpoint. That matters for any comparison: with the ElevenLabs engine selected in HeyGen, a HeyGen vs ElevenLabs test would compare ElevenLabs with itself. For this test I used HeyGen's own engine; in my voice list, the clone was labeled “HeyGen Voice”.

Chapter summary

An editorial summary, not a verbatim transcript.

00:00 Which one is really me?

The video opens with three clips of the same sentence: one is my real voice, one is the HeyGen clone, one is the ElevenLabs clone. You're asked to guess which clip is real; the answer comes later in the video. Both clones were made from the same recording of about two minutes.

00:40 Is HeyGen Voice really #1?

HeyGen says its new voice model is number one on a public leaderboard. That's true for Artificial Analysis' voice cloning list in English: on October 11, HeyGen Voice was at rank 1 to 2 there, sharing the top spot with Qwen. On the main leaderboard, ElevenLabs holds the top two spots, and HeyGen Voice isn't listed yet. Because the model is so new, I tested it myself. On screen: this video is not sponsored, I pay for both tools myself, I'm an independent ElevenLabs affiliate and I've applied to HeyGen's affiliate program.

01:13 How I tested both clones

Both clones come from the same recording: 2:13 minutes of free speech, with default settings. HeyGen let me pick one of three clone variants; ElevenLabs gives two takes per sentence, and I always took the first. The clone was ready after about 22 seconds in HeyGen and about 11 seconds in ElevenLabs. A tip for HeyGen: with a circle avatar look, generating failed for me without the error message saying why; with a full-frame look it worked.

01:56 Blind test 1: neutral sentence

The first blind test uses a neutral sentence: “Hey, it's Lukas. Today I'm testing whether a ten-second voice clone can really replace two hours of studio recordings.” Three clips, one of them is real. The reveal: number 1 was ElevenLabs, number 2 my real voice, number 3 HeyGen. To my ears, ElevenLabs was closer to my voice in this round.

03:27 Blind test 2: my name and a price

A sentence with tricky words: my name, Lukas Hüttis, and a price of 22 dollars, plus my German accent. Claude Code built the blind test from my files, so I didn't know the order myself. I correctly picked number 1 as the real me but mixed up the two clones: number 2 was HeyGen, and that's the one I found closer to my voice this time.

05:05 Blind test 3: tutorial line

A sentence in tutorial tone: “Open the settings, click on Voices and paste your script.” I was sure number 3 was me, and it was. I found number 2 closer to my voice, and that was ElevenLabs.

06:10 The intro clips, revealed

The reveal of the three clips from the start, with the series sentence “I built this entire project over a single weekend”: number 2 was me, number 1 HeyGen, number 3 ElevenLabs. Here I found HeyGen closer to my voice. After the four English rounds, that makes it 2:2.

06:49 Bonus round in German

German is my native language, so I'm strict here, and there's no German ranking for HeyGen Voice on the leaderboard. The sentence: “Hey, hier ist Lukas. Heute teste ich, ob ein Stimmklon aus zehn Sekunden wirklich zwei Stunden Studioaufnahmen ersetzen kann.” Number 3 was me, number 1 HeyGen, number 2 ElevenLabs. I found ElevenLabs closer, but not by as much as I'd expected. My hunch: you can hear whether a voice sounds like someone even if you don't speak the language.

08:30 Expression: tags vs no tags

This round is not blind. ElevenLabs understands audio tags like [laughs], [whispers] and [sighs] in the text; HeyGen doesn't. HeyGen reads the sentence without tags first, and it sounds pretty close to my voice. Then ElevenLabs with the tags. I love how ElevenLabs does the tags, and I'd wish for that in HeyGen as well.

09:41 Prices

As of October 11, billed monthly: HeyGen Creator costs $29 a month with voice cloning included. ElevenLabs Starter at $6 a month includes instant cloning, Creator at $22 a month adds professional cloning. When I recorded this, Eleven v4 was on a free two-week promotion. With HeyGen you also get the avatars, a huge bonus if you're focusing on that.

10:14 My verdict

For me it's basically a draw, 3:2 for ElevenLabs by my ears. These were first impressions with first takes only; with two, three or four tries per sentence you'd probably get a perfect version. I'd be happy with both: HeyGen's newest voice clone is now very similar to the newest ElevenLabs.

10:53 Free call and Part 2

If you want this for yourself or your business and don't know where to start, there's a link to a free call below the video. At the end I announce Part 2 of my ElevenLabs series: Eleven v4 as an instant versus a professional clone, plus the voice changer and more languages. Part 1 from July still ran on an older model version.

Related

Want a voice clone for your courses, explainer or product videos?

I work 1:1 with freelancers, consultants, coaches and small teams on practical AI workflows, automation and custom tools. The intro call is free and takes 15 minutes.

Book a free strategy call

More videos →