Synthesia 2

• 1. overview (724 chars)

  Synthesia is an enterprise-level AI digital human video platform headquartered in London. It was founded in 2017 by Victor Riparbelli and a group of AI researchers from UCL. Users input a text script to generate a "broadcast-style" video that is announced by a virtual anchor, lip-synced, and comes with multi-lingual dubbing. It is widely used in corporate training, internal communication, and product explanations. Product Express records it as "Synthesia 2", which is not an official single generation - the company's official current generation is Synthesia 3.0 (released in October 2025), and its underlying driver is the Express-2 full-body digital human engine launched in September 2025. This report is written based on Synthesia 3.0 / Express-2, the current verifiable mainline, and differences in naming standards are left for manual review.

• 2. Product background (802 chars)

  Synthesia is one of the few independent companies that has passed the "AI digital human video" business closed loop and has not been incorporated by major manufacturers. Its financing rhythm confirms the popularity of the track: Series D was completed in January 2025, with a valuation of US$2.1 billion; in January 2026, it received another US$200 million in Series E led by GV (Google Ventures), with the valuation doubled to US$4 billion. Investors also include NVIDIA, Accel, etc. Revenue is equally steep - annual recurring revenue (ARR) increased from $88 million at the end of 2024 to $146 million in September 2025. According to official and third-party data, more than 90% of the Fortune 100 companies are using the platform, and the cumulative number of services provided exceeds 60,000 companies. In April 2026, the company announced the establishment of offices in Austin, Berlin, Paris, and Zurich, and planned to expand its headcount by 70%.

• 3. Product functions and usage experience (1844 chars)

  Synthesia's core workflow is from "script to finished film": select a virtual anchor, write or generate scripts, configure scenes and dubbing, and render and export with one click. Two major upgrades in 2025 form the product skeleton today.   One is the Express-2 Digital Human Engine (September 2025). Different from the previous generation Express-1 (which will only make facial expressions in April 2024), Express-2 uses a diffusion Transformer to render the whole body - gestures, postures, and weight shifts all change naturally with semantics, with micro-expressions (winks, raised eyebrows, smiles) and stable eye contact, output at 1080p/30fps. In the actual test, the digital human will use gestures to emphasize points and change postures during transitions. At normal playback speed, it can basically cross the passing line of "resembling a real person". The second is Synthesia 3.0 (October 2025), which adds several key capabilities: Video Agents - virtual anchors who can be embedded anywhere in the video, can have real-time two-way dialogue with the audience, call the connected knowledge base (SharePoint, Google Drive, CRM) to answer questions, run training or initially screen candidates. It is currently launched in phases for enterprise customers; AI Playground - call Sora 2, Veo 3.1, FLUX.2 directly in the editor to generate a cinematic feel. B-roll material, no need to jump out of the workflow; as well as image/attire generation based on prompt words, and Personal Avatar that "generates a personal digital avatar from a single photo".   On the dubbing side, Express-Voice is a two-stage Transformer voice cloning system (800 million parameters for each stage), which can reproduce the timbre in seconds and preserve the accent and intonation rhythm. AI Dubbing can translate a video into 130+ languages ​​and maintain frame-level mouth alignment, while one-click translation can perform script-level translation. The platform has built-in 240+ digital people in stock, supports text-to-speech in 160+ languages, and provides real-time collaboration, Brandkit brand suite, version control, SCORM export (connected to LMS) and playback analysis.

• 4. Pricing and business model (722 chars)

  Pricing in April 2026 adopts a "minute quota" annual payment system: the free version is 36 minutes per year, 9 stock images (with watermark); Starter is $29 per month ($18 annually), 120 minutes, 70+ images per year; Creator is $89 per month ($64 annually), 360 minutes per year, unlocks full inventory; Enterprise is a customized quote, including SSO, audit logs, SCORM, API, Veo 3 and more. Personal Avatar (personal digital avatar) is additional, about $1,000/year. It should be reminded that SCORM export and one-click translation have been criticized by many reviews as being "exclusive for the enterprise version" or requiring sales negotiations to be activated. There is often a gap of "building a course only to find that it cannot be used". The API has been devolved from pure Enterprise Edition to Creator Edition for programmatic batch generation.

• 5. User reviews (874 chars)

  The positive points focus on its ease of use and multi-lingual capabilities: people with zero video experience can produce a video within 15 minutes, and 90% of users can publish their first video without tutorials; an English training video shot in one shot can be quickly translated into dozens of language versions, which is of direct value to global L&D teams. Negatives are highly concentrated on content review: this is the most common complaint - compliant commercial content will be intercepted or blocked without clear reasons, manual review often takes 12-24 hours, and non-enterprise customers have no priority; some users reported that their accounts were blocked within 30 minutes due to vague violations. Strict refund policies and slow custom image production (it takes several days to review) are also common shortcomings. Other users reported that rendering would take 15-20 minutes during peak hours, and that the monthly 10-minute Starter quota could easily reach the top, forcing them to upgrade at 4 times the price.

• 6. Industry evaluation (785 chars)

  In independent reviews, Synthesia's overall reputation is positive: G2 is about 4.7/5 (2000+ reviews), Trustpilot is about 4.3/5 (300+), and PlainAI is 4.0/5. The industry generally regards it as the standard answer for "corporate training/internal communication videos". Its rival HeyGen is more advantageous in terms of low-threshold realism, 175+ language translation depth, pay-as-you-go API and cost-effectiveness, but its governance and review are looser. General-purpose video generation models such as Runway, Sora, and Kling are squeezing the purely digital human broadcast category from the "movie-like B-roll" side; Colossyan has its own niche in native interaction in training scenes, and D-ID has its own slots in API-based photo drivers. Synthesia's moat is compliance credentials (SOC 2 Type II, GDPR, ISO/IEC 42001 AI Management System, SAML/SSO) and the purchasing inertia of 90%+ of the Fortune 100.

• 7. Controversies and Risks (823 chars)

  The biggest risk is the unpredictability of content review: users in regulated industries (medical, financial, legal) may have their approved videos re-labeled due to a slight edit, and the work hours invested are wasted; the platform's official documents admit that the review is not perfect, but do not disclose a clear appeal process, and non-enterprise customers have no priority channel. The second is the compliance gap - as of March 2026, neither Synthesia nor HeyGen has disclosed a HIPAA Business Associate Agreement (BAA), and regulated buyers involving patient or customer data need to request it separately before signing. Once again, there is the ceiling of realism: 64% of the audience in the blind test in 2026 can correctly identify the AI ​​digital human (89% in 2024). It still looks synthetic in close-up and is not suitable for emotional narratives or cinematic content. HIPAA, strict refunds, tight minute quotas, and rendering delays all create friction.

• 8. Applicable groups and usage suggestions (720 chars)

  Synthesia is best suited for: Video teams that need scalable, scalable, multilingual enterprise training and internal communications; Organizations that value compliance credentials and LMS integration. It's not suitable for: Creators looking for personality, humor, or a cinematic quality; Individuals on an extremely tight budget who only produce occasional films (the free version or cheaper tools are more cost-effective); Teams who need HIPAA-compliant medical content (BAA needs to be confirmed first). If you mainly do external marketing, HeyGen is often more cost-effective; if you do branch-type interactive training, Colossyan's native interaction is stronger. It is recommended to use the free version to produce a real business video first, and then consider paying after confirming that it can pass the review and the image fits.

• 9. Summarize (299 chars)

  Synthesia 3.0 / Express-2 has achieved an enterprise-level level of "typing to generate broadcast videos". Its real strengths lie in compliance, scale and multilingual training videos, while its shortcomings lie in the uncertainty of review, the ceiling of realism and the pressure of billing by the minute - it is an expert tool, not a universal machine.

No in-depth report yet

• 

  Lilly694

  —

  口型近看还是有点假，客户问过两次。

---

• 

  CHwhi

  —

  和 HeyGen 比，Synthesia 在企业合规、LMS 集成上稳得多；但写实度和多语种深度 HeyGen 更强，看你用在哪。

---

• 

  Henry.GonzalesJr

  —

  express-2 之后手势和姿态终于像真人了，强调处会用手比划，正常速度看基本过关，近景还是能看出是合成的。

---

• 

  IsaiahNgyyen

  —

  Trustpilot 上 4.3 分在 AI 工具里算异常高，负面基本集中在审核和退款，正常用其实挺稳的。

---

• 

  KathleenMiller_Pro03

  —

  零视频基础，二十分钟就出片，太顺手了。

---

• 

  Sean_Flores369

  —

  退款政策挺严的，短期订阅基本不退。

---

• 

  唐琪_1

  —

  免费的 10 分钟额度，一条就没了，坑。

---

• 

  CCastilloQ03

  —

  AI Playground 里直接调 Sora 2 和 Veo 3.1 生成 B-roll 是真香，不用跳出工作流也不另付 API 费，这是和竞品拉开差距的地方。但没有时间轴编辑，想逐帧微调根本做不到，改个逗号都要重新渲染等八到十二分钟，重度迭代会很磨人。

---

• 

  AMoore_66

  —

  数字人库是见过最全的，我居然找到一个长得像我的。就是 Starter 的分钟限额太别扭，后来升了档才好用。

---

• 

  JRichardson_20230

  —

  语音克隆几秒出声，口音语调都保住了，牛。

---

• 

  GloriaReev

  —

  我们团队把 40 页的 PPT 一天之内转成了完整视频课程，它居然会读演讲者备注生成配音，离谱地好用。不过长视频渲染是真慢，有次五分钟以上的片子直接导出超时，后来我们都习惯先把文案锁死再点生成，省事不少。

---

• 

  DylanPerez_2022

  —

  Personal Avatar 现在一张照片就能做，以前要录几分钟视频，速度提升巨大，做数字分身终于不痛苦了。

---

• 

  DianeCollins_2024

  —

  注册一小时就做出第一条视频了，界面真清爽。

---

• 

  ANjoh

  —

  多语种译制是真的救命，一个英文培训片直接翻成德文法文，省掉三次重录。质量大概九成到位，就是德文版偶尔口型比声音慢半拍，客户凑近看会问，但整体还是甩同行一条街，尤其对全球 L&D 团队，长期看特别值。

---

• 

  MariaSanchez

  —

  盲测里 64% 的人能认出是 AI 数字人，比 2024 年的 89% 低了不少，说明写实度确实在涨，express-2 之后手势姿态像真人了。语音克隆也狠，盲测 72% 分不清真人，但真要做情感叙事或电影感内容它还是不行，那是另一类工具的事。

---

• 

  LambdaLPWood

  —

  我们做全球入职培训，同一门课要存在六七种语言里。以前每上个市场重录一次，现在上传英文版一键译制，两小时搞定。省下的真金白银比订阅费多太多了。就是自定义形象要走同意流程，不是秒出，运营上有一点开销，但可接受。

---

• 

  DeborahCarter

  —

  Starter 那 10 分钟每月额度纯属玩笑，一个片子就烧光，逼着你升 Creator，实际月费直接翻四倍，签约前真不会告诉你。

---

• 

  JReed_Pro

  —

  高峰时段渲染能拖到十五二十分钟，长视频更夸张。我们后来学乖了，先把文案锁死再生成，绝不生成后再改逗号。

---

• 

  Lawrence.CampbellZ

  —

  内容审核是我最想吐槽的：正经商业内容会被无理由拦下来，人工复核动辄 12 到 24 小时，非企业客户还没优先级。而且近景下数字人还是能看出是合成的，要是你的流程经不起这突如其来的延期，这是个真风险，得留缓冲。

---

• 

  HashNet_pro

  —

  SCORM 导出居然是企业版专属，建完课程才发现，得找销售要报价，体验像钓鱼。

---

• 

  TIram

  —

  做 HR 合规课每季度都要更新，以前拍片子要折腾一整天，现在改脚本重新生成数字人视频一小时就搞定，企业版是贵但替换掉的东西值这个价。不过做对外营销我还是更偏向 HeyGen，翻译和写实度更顺，Synthesia 强在内部培训和 LMS 集成。

No reviews yet