HeyGen has made its Voice Turbo speech model free to use until 30 November 2026. In an announcement updated on 2 November, the company describes Turbo as a faster version of its own voice model: a user supplies text and selects a voice, and the service generates a spoken recording. A public voice is enough to try it; recording yourself is optional.
The model and the voice are separate choices
The HeyGen Voice documentation lists standard Voice 1, the faster Turbo and Turbo Free. The latter two use public voices and instant clones from a recording. Professional cloning requires standard Voice 1. The voice identifies the selected speaker; the model determines the available synthesis mode and speech controls.
| Model | Available voices | Terms |
|---|---|---|
| Voice 1 | Public, instant and professional | Paid synthesis |
| Voice Turbo | Public and instant | Faster paid mode |
| Voice Turbo Free | Public and instant | Free offer until 30 November |
A comparison of recordings therefore needs to identify both the model and the voice. Changing the speaker changes the impression of timbre and delivery even when the synthesis system remains the same. Keeping the reference voice and text fixed makes a model comparison more useful.
A short recording or a professional clone
The Instant Voice Clone mode creates a voice from one recording. It uses the first three minutes and, according to the developer, is usually ready within seconds. An instant voice cannot be retrained: a new recording creates another voice. This offers a way to try a personal speaking style without preparing a long studio session.
For a Professional Voice Clone, the service requires one to ten recordings of the same speaker, totalling at least 20 minutes. It trains a dedicated adapter—a set of settings for that voice—and the voice becomes available after training completes. A professional voice occupies a paid slot in the account and can be retrained with new recordings. The developer recommends removing extended silence, background noise and overlapping speakers from the source material.
Expressive controls depend on the voice type
Public and instant voices understand HeyGen audio tags: directions for emotion or whispering, and sounds such as laughter, a sigh or a pause. The developer distinguishes a change to subsequent delivery from a sound placed at a particular moment. A delivery tag affects up to approximately 17 seconds; a longer emotional passage needs repeated directions. Professional voices use a different set of controls.
The speech-generation documentation gives professional clones a speed range of 0.75 to 1.25, pitch adjustment within four semitones and control over pitch variation. Instant voices, including Turbo, accept these professional parameters but ignore them.
What an independent comparison measures
On 9 November, the Artificial Analysis Controlled Voice Arena gives HeyGen Voice 1189 ± 13 Elo and Turbo 1173 ± 16. Qwen-Audio-3.1-TTS-Plus sits between them at 1179 ± 13. The confidence intervals overlap, and the order of these entries can change.
Under the Artificial Analysis methodology, models read the same texts using eight cloned English voices: four US and four UK. Listeners choose between two recordings without seeing the model names. This reduces the influence of a preference for an individual speaker. Elo expresses relative listener preference, rather than the percentage of words pronounced correctly.
That comparison concerns English. It cannot establish Russian stress accuracy or the reading quality of a particular technical article. Artificial Analysis uses a separate test for challenging pronunciation sequences; an appealing delivery and the exact reading of a number or name are different properties of speech.
Russian and longer texts
The HeyGen speech API parameters list 39 language codes, including Russian and English. A request can contain up to 5,000 characters. The service can return a completed recording or deliver audio while it is generated; a longer article continues through multiple requests. This is a limit on an individual passage, rather than the total duration of an audio edition.
You can also hear the model on pommeDeTerre: our article audio editions are generated with HeyGen Voice Turbo Free. Russian texts use a saved Russian voice; English texts use a standard public voice.
The November 8 AI news overview offers its complete audio edition in the player above the text. It offers a real editorial text containing model names, numbers and extended explanations.
Free access depends on shared capacity
The API usage limits allow 30 speech requests per minute per workspace member, counted separately for each speech endpoint. Turbo Free also shares limited computing capacity: a request can be rejected even when the user has not exhausted that allowance. Availability and waiting times are not guaranteed for the free offer.
HeyGen says it will announce in advance how access will change after 30 November 2026.
