Source: raw/Best_AI_Voice_Generator_-_HeyGen_vs_ElevenLabs_vs_Cartesia_vs_Inworld_vs_Fish.md (youtube.com/watch?v=Vnl0IkAwdaE, uploaded 2026-10-28, 10:29)
Creator: HeyGen official YouTube channel (@heygen_official); presenter not named in the transcript.
URL: https://www.youtube.com/watch?v=Vnl0IkAwdaE
Duration: 10:29
Platform: YouTube
Skill level: Beginner
A presenter on HeyGen’s own channel clones their voice from one recording on five models from five companies, has each read the same unseen script, and scores them 0-5 in six categories. HeyGen Voice wins with 28.5/30. Treat this as a vendor-made comparison: HeyGen published it, the winner is HeyGen’s product, the scores are one person’s subjective listening, and the video closes with “subscribe to HeyGen.” The method is the most reusable part, and it doubles as a template for testing a voice clone yourself.
Key Takeaways
- Final ranking (presenter’s subjective scores out of 30, HeyGen-made test):
| Rank | Model | Score | Presenter’s one-line verdict |
|---|---|---|---|
| 1 | HeyGen Voice | 28.5 | Closest match to the original vocal identity; most consistent; slightly less emotive than ElevenLabs |
| 2 | ElevenLabs v4 Turbo | 27.5 | Strongest emotion; very natural, but “slightly more polished or performative” than the original |
| 3 | Cartesia Sonic 3.6 | 26.5 | Clean and balanced; restrained emotion, which also keeps it from drifting off the speaker |
| 4 | Inworld Realtime TTS-2 | 25 in the recap (20 in its own segment) ^[ambiguous] | Fastest read (~18.22 s); brighter than the original voice; occasionally rushed |
| 5 | Fish | 24.5 | Good identity preservation; slowest read (~23.93 s), so phrases feel drawn out |
- Category winners:
- Similarity: HeyGen closest; Cartesia and Fish also strong.
- Naturalness: HeyGen, ElevenLabs and Cartesia tie at 4.5.
- Emotion: ElevenLabs v4 Turbo “is the clear winner.”
- Pacing: HeyGen, ElevenLabs and Cartesia tie at 4.5; Inworld feels rushed, Fish drawn out.
- Pronunciation: HeyGen and ElevenLabs tie with perfect scores.
- Consistency: HeyGen highest.
- Pick by priority, not by total (presenter’s own conclusion): identity match → HeyGen Voice; strongest emotional delivery → ElevenLabs v4 Turbo; clean, undramatic delivery → Cartesia Sonic 3.6. Inworld produced the shortest result but sometimes felt rushed; Fish kept the identity but its slow pacing made it the weakest fit for this script “without further adjustments.”
- The lesson worth keeping: a realistic AI voice is more than a cloned timbre. One model keeps the identity but struggles with emotion, another sounds human but drifts from the speaker, a third pronounces everything right but paces it differently. So test a clone on new text it hasn’t heard, in the type of content you actually plan to make, rather than choosing from a leaderboard or a polished demo.
- How the models were chosen: starting from the Artificial Analysis Controlled Voice Leaderboard. At recording time the presenter reports: HeyGen Voice Turbo #1, HeyGen Voice #2, Qwen Audio 3.1 TTS Plus #3 (captioned “Kwen”), ElevenLabs v4 Turbo #4, Cartesia Sonic 3.6 #6, Inworld Realtime TTS-2 #7. Rank #5 isn’t named.
- Excluded: Qwen, because it “requires a more technical API workflow,” and HeyGen Voice Turbo, to keep one model per company.
- Speechify AI Simba 3.2 ranked above Fish and was planned for the test. Speechify showed voice cloning as available and prompted an upgrade; after the presenter paid, cloning was greyed out as unavailable. The presenter says this “should have been disclosed before checkout.” Fish replaced it.
- Relation to the rest of this topic: the emotion verdict matches the existing view that ElevenLabs leads on emotional range (Voice Agent Comparison). Cartesia’s model is now at Sonic 3.6, past the Sonic 3.5 covered in Cartesia Sonic. This test is about cloned-voice quality for pre-rendered video; it says nothing about latency or real-time voice agents.
Test Method (reusable)
- One source recording for every model. It was deliberately hard: a time (“9:47 this morning”), a person’s name (“Dr. Maya Chen”), a version number (“version 3.1”), a foreign place name (“Sao Paulo”) and a deadline (“by Friday”).
- A different, unseen test script, identical word for word and punctuation for punctuation across models. It includes a question (“Is it perfect? Not quite.”) and a closing statement. This tests whether the clone generalizes rather than replaying a performance it has heard.
- No provider-specific delivery prompts. Each model got neutral, identical conditions. (Fish was generated directly through Fish’s own platform.)
- Six categories, 0-5 each, half-point steps:
- Similarity: resemblance to the original recording.
- Naturalness: sounds like a real person.
- Emotion: feeling and emphasis.
- Pacing: speed, pauses, rhythm.
- Pronunciation: clarity and accuracy.
- Consistency: identity and quality stay stable across the sample.
- Listen against the original. Each sample was played without voice-over and compared with the source recording. The presenter says plainly that these are subjective scores, “not official benchmark scores.”
- Measure read length. Total duration (Inworld ~18.22 s, Fish ~23.93 s) exposed pacing differences that the per-sentence listen didn’t make obvious.
Try It
- Before committing to a voice vendor for a client or channel, copy this method: record one sample with numbers, names and a foreign word in it, then write a second script in the content style you’ll actually ship (a dental explainer, an ad read, a lesson).
- Clone on your top 2-3 candidates from that one recording, generate the unseen script with no special prompting, and score the six categories blind, ideally with someone other than the speaker also listening.
- Weight the categories by use case. For a personal-brand avatar, similarity and consistency matter most; for an ad or story read, emotion and pacing matter more.
- Record the read length of each output; a long or short outlier usually means a pacing problem you’ll need to fix with prompting or editing.
- Check that voice cloning is actually included on a plan before paying. The Speechify episode shows a feature can be advertised and then unavailable after checkout.
Open Questions
- Inworld’s score is inconsistent in the source. Its own segment says “20 out of 30”; the final recap says it “finishes fourth with 25.” The order of 4th and 5th place depends on which is right (Fish has 24.5). Per-category scores for each model weren’t read out in full, so the total can’t be rechecked.
- Vendor bias is built in. HeyGen produced and published the test, HeyGen Voice was tested first, the single scorer is presenting on HeyGen’s channel, and the source recording was the presenter’s own voice. No independent replication is in the vault.
- Which Fish model was tested is not stated (only “Fish”). Voice Agent Comparison covers Fish Audio Drama 3, but it is unknown whether that was the version used here.
- Leaderboard naming: the presenter cites the Artificial Analysis Controlled Voice Leaderboard; Cartesia Sonic’s sources cite the Artificial Analysis Text to Speech leaderboard. Whether these are separate boards wasn’t checked. The rankings above are as of the video’s recording date, not verified against the live board.
- Model names from auto-captions: “Kwen Audio 3.1 TTS Plus” is read here as Qwen; “11V4 Turbo” / “Eleven V4 Turbo” as ElevenLabs v4 Turbo. Neither was checked against vendor docs.
- Pricing, latency and API availability of each model are not covered.
Related
- HeyGen Voice — the winning model: scene-adaptive cloning, expressiveness slider, speech-to-speech, and HeyGen’s own preference data
- Cartesia Sonic — Sonic 3.5 background, latency claims and pricing; this test used Sonic 3.6
- Voice Agents with Claude Code + ElevenLabs — the ElevenLabs voice-clone and agent workflow
- Voice Agent Comparison — architecture, latency and pricing view, including the October 2026 Fish Audio Drama 3 and ElevenLabs v4 notes
- HeyGen Avatar V — the avatar model HeyGen Voice drives
- Fluid Voice and Voice Box — free self-hosted cloning, the option outside this paid-vendor field
- AI Avatar Generators Compared (2026) — the companion HeyGen-made test for avatar tools
- AI Video Generators Compared (HeyGen-made test, 2026) — the companion HeyGen-made test for video generators