HeyGen

What HeyGen is

HeyGen is a video production platform with avatars, voices and audiovisual localisation. Its current offering combines Avatar V, Video Agent and the AI Studio editor, alongside tools for adapting existing recordings. It extends beyond animating a photograph.

Distinguish three outcomes: a synthetic presenter reading a script, a dubbed recording and an assembled piece containing graphics and imagery. Each needs different materials and review. A team may use the platform for all three while retaining a record of what came from authentic recording and what was generated or transformed. This distinction helps explain the finished piece to its audience.

Reference documentation: Avatar V product · Video Agent.

From Movio to HeyGen

Joshua Xu and Wayne Liang developed the project known as Movio, which became HeyGen as its scope expanded from spokesperson videos to broader audiovisual creation. A $60 million Benchmark-led round was announced in June 2024.

Historical company pages have been updated since the events: an article’s revision date should not be confused with the date of financing or rebranding. Product development also requires care when reading older guides. Instant Avatar, Digital Twin and Personal Avatar appear in documentation from different stages, while actual capabilities depend on the current model and workflow. Record these identifiers when planning a repeatable production process.

Reference documentation: Movio rebranding announcement · Series A announcement · Forbes contemporaneous financing report.

Avatar V: visual reference and voice

Avatar V uses a reference recording; Avatar IV starts from a photograph. Research published in April 2026 describes a diffusion transformer with flow matching, conditioned on video and audio, and progressive training across five stages.

The model aims to retain identity and reproduce performance connected to the voice. For users, source material matters: lighting, framing, posture and sound clarity determine the available reference. A sample should not be selected merely because it is flattering. It should represent the speaking and movement style intended for the series, without gestures that distract from the message. Assess the resulting performance across several sentences rather than a single still frame.

Reference documentation: Avatar V product · Avatar V research,8April2026.

Video Agent, editing and reuse

Video Agent prepares pieces from instructions, combining a presenter, imagery and graphics. Motion elements can be edited in AI Studio. AI Clipping selects excerpts from recordings and supports XML export for continued editing elsewhere.

An automatic draft can suggest structure, but the objective should supply the editorial criteria. To turn a long session into short pieces, identify one complete idea per excerpt and retain the conditions giving it meaning. Pay particular attention to cuts removing a negation, exception or the question that prompted an answer. An appeal score cannot replace this check, especially when the excerpt will circulate without the original context.

Reference documentation: Video Agent · AI Clipping.

Presentations and documents as input

PPT or PDF import offers an editable template or slides as background images. Each script segment is linked to its scene. This choice determines how much can subsequently be changed without rebuilding the design.

A background image preserves appearance but does not make its text editable. For multilingual presentations, decide whether to retain the original design or rebuild titles and diagrams. Review tables, figures and cropping after import. Reserve space for the avatar without covering information. A slide readable during an in-person presentation may need simplification for a phone screen, where labels and the presenter compete for a smaller area.

Reference documentation: PPT/PDF to Video.

Synthetic voice and performance direction

HeyGen Voice 1 provides text-to-speech and speech-to-speech conversion. In Studio, Voice Mirroring uses a recorded performance as its basis. The offering distinguishes instant cloning, professional cloning and a Turbo API variant.

The choice depends on the control needed over delivery. Text lets you change words; a reference take communicates pauses and emphasis directly. For a short instruction, record the condition and action clearly, then listen for preservation of both. Keep pronunciation correction separate from content changes. A different voice can improve intelligibility without justifying alteration of approved meaning, and a technically clean recording still needs listening review.

Reference documentation: HeyGen Voice1.

Translation: engines, timing and sound

Translation distinguishes Audio Only, Speed and Precision; the latter two include lip synchronisation. It requires a single source language. Dynamic Duration can change timing, and sound effects are removed from translated audio even when music is retained.

Choose according to the material. A stable, front-facing speaker differs from a conversation with changing shots and partially obscured faces. Sound can carry meaning in a demonstration; do not automatically treat it as decorative noise. Compare the original and adaptation, recording timing changes and any acoustic cue that needs reinstating through reviewed editing. Check the final mix rather than assuming that successful speech translation preserves the whole soundtrack.

Reference documentation: HeyGen — Video Translation.

Glossary and correction before generation

Brand Glossary guides terminology and pronunciation. Proofread, also called Review & Edit, supports review of the translated script before final production. These tools provide editorial control within localisation.

Prepare a small set of troublesome terms: acronyms, equipment names and internal labels. Include enough context to distinguish words whose meanings change. Review a sentence containing a condition and exception first, then check numbers, units and recipients. When a translation changes, record who accepted the correction and verify that the final audio uses the intended wording. A glossary is most useful when it reflects actual decisions rather than becoming an unreviewed list of suggested translations.

Reference documentation: Brand Glossary · Proofreading for Video Translation.

Visible and switchable captions

AI Studio supports styled captions incorporated into the image. The guide distinguishes these from switchable captions on shared translation-project pages; avatar videos use the styled, rendered option.

Choose according to the channel. Silent playback needs visible information; a player with separate tracks gives viewers a choice. Avoid covering signs, diagrams or buttons. Check contrast and size on mobile and allow sufficient reading time. Words embedded in screenshots or slides need separate review even when transcription and voice are correctly translated. Test the delivered file in its actual player because an editor preview does not establish how captions will appear after distribution.

Reference documentation: HeyGen — Captions.

Export, colour and SCORM

Delivery includes MP4 and SRT, with resolution depending on plan and source material. HeyGen processes video in SDR and does not provide HDR output. Business and Enterprise support SCORM 1.2 or 2004 and watched-percentage completion criteria; Auto-SCORM keeps video hosted by HeyGen.

Choose between a fixed file and a service-linked delivery. The latter simplifies updates but needs control over regeneration of a version already used by learners. Test playback, completion status and campus access. For footage from an HDR phone, review colour and contrast after conversion; a visual change may affect the readability of labels or screen captures.

Reference documentation: PPT/PDF to Video · Video download and export · SCORM export.

LiveAvatar and real-time conversation

LiveAvatar is a separate platform for real-time conversation, using WebRTC and API or SDK integration. Its avatars are not directly interchangeable with HeyGen avatars, and its credits and plans are independent.

A video follows a closed script; a conversation can receive unexpected questions. This changes the project. For guidance on internal information, define sources, supported questions and a route to a person. An expressive interface is not automatic technical authority. Communication practice should use fictional situations and clear criteria, including tests of what happens when someone asks a question outside the intended scope. Evaluate the response system separately from visual realism.

Reference documentation: Introducing LiveAvatar · LiveAvatar FAQ.

Platforms, plans and collaboration

Plans include Free, Creator, Pro, Business and Enterprise, with different credits and model access. The mobile application is more limited: documentation directs advanced AI Studio functions to desktop and states that workspace switching is unavailable on mobile.

Before producing a series, establish where editing will happen and who will make corrections. A mobile trial does not demonstrate that the entire team workflow is supported there. Estimate needs per accepted production: minutes, languages, retries and review. Keep approved versions in a designated folder and separate personal experiments from material representing the organisation. This also makes responsibility clearer when an old video needs withdrawal.

Reference documentation: HeyGen Mobile · Plans.

Identity, consent and data

The represented person must record their own consent even when someone else organises the project; a link or QR code can support this. Policy permits requests to remove likeness. HeyGen excludes Enterprise data from training by default; other users can request exclusion. It states that data is stored in the United States.

The agreement should cover messages and distribution as well as avatar creation. Define what happens if the person changes role or withdraws permission. Remove unnecessary customer or worker data from references. A prevention video does not need an individual’s health history. Retain permission records and an inventory of pieces using the person’s face and voice.

Reference documentation: Security Practices · Privacy Policy,15October2026 · Consent recording · Moderation Policy.

Reception and impersonation risks

Group-IB has noted malicious actors’ interest in bypassing controls in avatar applications such as HeyGen. Provider policy prohibits fraud, disinformation and deceptive impersonation. The existence of controls does not establish that every circulating piece has been authorised.

In corporate communication, the risk is not limited to a false face: a familiar voice can make an unapproved instruction believable. Publish through identifiable channels, include responsibility and context, and retain the authorised original. If a received video is attributed to someone, verify the instruction through the usual channel before acting on sensitive changes. The communication process remains important even when the media looks convincing.

Reference documentation: Group-IB deepfake overview · Moderation Policy.

Practical applications in OSH

Its potential occupational safety and health role is communicating general instructions and induction messages to diverse audiences. It can help produce versions of approved content, provided reviewers check that each translation preserves the conditions, exceptions and actions established by the original message. Review should include the finished video, not just the script.

It can adapt internal communication for audiences using different languages, maintain short explanations or update induction videos. In OSH, consistency between narration, wording and the actual action channel matters more than making every language version exactly the same length.

A useful translation preserves meaning and enables action. If a literal expression sounds unnatural in the target language, review it linguistically without changing the approved instruction. Keep a record of agreed terminology so later videos do not introduce conflicting labels.

Worked example and result checks

Example: prepare a short video about an internal hazard reporting channel in Spanish, Catalan and English. The team reviews all three versions, confirms pronunciation of names and locations and checks that on-screen text matches the narration and actual reporting process.

1. Validate the source script and prepare a glossary of names, acronyms and terms requiring consistency.
2. Translate a representative segment containing an instruction, exception and contact. Review it before processing the full video.
3. Compare audio, captions and visible wording in each language. Correct titles that remain in the source language.
4. Listen to the complete result for ambiguous pronunciation or pauses changing the meaning.
5. Check that someone can identify when to use the channel and how to access it. Retain versions and withdraw obsolete ones when details change.

This is a proposed localisation workflow. It does not claim that a particular video has been reviewed by a competent speaker or understood by every viewer.

Limitations and conditions of use

Visual synchronisation and fluent speech do not guarantee an accurate translation. Review technical terms, negations and proper names; use personal images or voices only with authorisation. An avatar must not imply that a real professional issued a message they have not reviewed.

The source recording affects the result: multiple speakers, frequent cuts and noise can complicate interpretation. Do not assume the quickest option suits a sensitive explanation. Check which correction tools the account provides before planning many versions.

Identity and voice require appropriate permission. Avatars or synthetic voices should not attribute unauthorised statements to someone. Clearly identify responsibility for the content and which elements use audiovisual generation.

Reference documentation: HeyGen — Video Translation · HeyGen — Captions.

When to choose it and how to complement it

HeyGen fits priorities around audiovisual localisation and version review. Synthesia provides another script-and-avatar workflow to compare. DeepL can help prepare text, with subsequent review of meaning.

The selection trial should include a correction after translation: check whether it updates audio, captions and visible wording or whether each needs separate work. A shared version list helps identify inconsistencies before release and assigns responsibility for maintaining each language.

Frequently asked questions

Does fluent speech guarantee correct translation?

No. Check terminology, conditions and meaning, particularly negatives and exceptions.

Are screenshots automatically translated?

Audio translation does not itself change wording embedded in an image. Review that content separately.

Should every language be generated at once?

Validate a sample and review method first to avoid multiplying an error in the original script.