Run a 30-Second A/B Before Swapping to HeyGen Voice

If your dubbing pipeline is running an incumbent TTS model, a new leaderboard rank is a starting point, not a green light. HeyGen shipped its first in-house text-to-speech model, HeyGen Voice, on October 9, 2026. In Artificial Analysis' Controlled Voice TTS Arena, HeyGen Voice had 1,201 Elo over 1,468 appearances, beating Qwen-Audio-3.1-TTS-Plus at 1,182, Eleven v4 Turbo at 1,166, and Eleven v4 at 1,162.

The arena gap is the entry fee

A few points behind the leader is not enough to migrate. A large gap gives you a reason to run a sample. The leaderboard tells you the voice wins head-to-head comparisons. It does not tell you whether your product names, dates, and brand terms survive your script. A small lead is easy to flip with your own content. A larger lead is more persuasive, but still not a guarantee. The arena measures perceived quality in the abstract; your dubbing job is narrower: it has to sound right on the exact words you ship.

Run the A/B before you touch the pipeline

The swap test has two checks: first, compare the arena Elo, pronunciation robustness, and price per 1M characters; second, A/B a 30-second sample of your actual script against your incumbent model. The checklist below covers those areas. Each item is quick.

• Pull the arena Elo for HeyGen Voice and your current model, then note the gap.
• Generate the same 30-second script through both voices and mark every mispronounced name, date, or acronym.
• Calculate your monthly character volume against the per-1M-character rate.
• For a clone, confirm the consent flow and training window before you record anything.

Use a script that includes your hardest words: product names, dates, abbreviations, and any non-English terms. Do not use a polished marketing line if your real pipeline is technical. The sample should be short enough to listen to again, but long enough to expose a pattern.

Keep the audio files side by side. Do not judge the new voice in isolation. Note any known flaw in the old voice. A fix in the new voice is a concrete reason to move.

Put the sample, the Elo gap, and the cost line in one note. That note is the decision record when someone asks why the pipeline changed.

The pronunciation number is the stress test for brand terms. HeyGen Voice scored 83.1% on Artificial Analysis' pronunciation robustness test and ranked 10th out of 29 models. That is strong, but not invincible. Your narration may lean on brand names, model numbers, or non-English names. The sample will show what the benchmark cannot.

The price math is where the swap lives or dies

HeyGen Voice is free inside HeyGen's own platform and API. The reported list price is $30 for every 1 million characters, with generation speed around 40 characters per second. On OpenRouter, HeyGen Voice carries a $15 per million character rate, while HeyGen Voice Turbo carries $7.50 per million, with both discounted 50% until October 31, 2026.

Write the monthly character count next to the rate. Multiply it by the per-1M-character price, then compare the monthly result to your current TTS cost. For small batches, the list price may be trivial; for scale, the per-character rate and discount deadline decide. Speed matters, because a slow queue can make a cheap voice feel expensive. Decide whether queue time is a real cost. Check the exact endpoint you will call, because a faster variant can change the math.

Voice cloning adds a consent line to the checklist

HeyGen sold a $99-per-month Professional Voice Clone add-on, trained with the speaker's explicit consent using 30 minutes to three hours of their speech. For a small business, a monthly fee is a different decision from a per-character API bill, and it is a consent decision. A client, employee, or creator voice makes the consent flow part of the A/B, not an afterthought.

Treat a clone as a permission record, not just a voice file. Keep the consent record and the training duration in the same folder as the sample audio. When the speaker leaves, or the brand changes, that record is what you show.