HeyGen Releases Three Builder Tools for Real-Time Avatars, AI Video Evaluation and API Voice Cloning

HeyGen said it shipped three developer-focused tools in September: an open-source LiveAvatar integration with OpenAI's GPT-Live-1, a Kaggle-hosted Code2Video Benchmark, and API access to Professional Voice Clone. The releases point to a broader push to make AI video less like a closed creative app and more like an agent-ready software stack.\n\nThe real-time avatar release is a public reference implementation that connects OpenAI GPT-Live with HeyGen LiveAvatar. The GitHub project describes GPT-Live as a full-duplex speech-to-speech model driving a real-time avatar whose tool calls appear as animated overlays. The demo is intentionally minimal: it uses a server orchestrator, browser client, LiveKit, HeyGen LiveAvatar and GPT-Live, with the browser kept away from API keys. It requires Node 20.12 or later, pnpm, a LiveAvatar API key and an OpenAI API key with GPT-Live access.\n\nThe timing matters because OpenAI recently brought GPT-Live-1 to the API, describing it as a full-duplex voice model for real-time conversations, priced at $0.05 per minute for the front-end voice layer. For avatar applications, that reduces the need to stitch together separate speech recognition, language-model and text-to-speech systems. Developers are treating that as useful for tutoring, live presenters, sales training, customer support and interactive streaming, while also noting that real-time systems still require careful handling of microphone permissions, API access, turn behavior and production hardening.\n\nHeyGen's second release is the Code2Video Benchmark, created with Kaggle to test whether AI agents can generate motion-graphics videos as code rather than raw pixels. The benchmark includes 168 curated, human-created reference motion-graphic videos across product-launch beats such as hook, problem, product intro, key feature, benefits, social proof, call to action and brand outro. HeyGen says model outputs are rendered through HyperFrames and then evaluated in round-robin pairwise comparisons against human references and competing models using a judge model.\n\nThe benchmark's results underline how hard visual craft remains. HeyGen reports that the top four models sit within 12 Elo of one another, open-weight models are within 50 Elo of first place, and models are closer on prompt intent than on composition and craft. Kaggle is hosting the Code2Video leaderboard, with runs executed on the platform so traces and logs can remain auditable.\n\nThe third release, Professional Voice Clone, is a paid API feature that trains a dedicated HeyGen Voice adapter from at least 20 minutes of recordings by one speaker. The API accepts 1 to 10 recordings, creates a voice through POST /v3/models/audio/voices, returns 202 Accepted while training runs, and requires polling until the voice becomes ACTIVE. Each professional clone uses a purchased voice slot, with five pooled trainings per monthly billing period per slot, including the initial training. Once active, the voice can generate completed WAV output or streaming speech.\n\nThe package is likely to appeal to teams building conversational video products rather than one-off clips: the avatar stack handles live presence, the benchmark gives model builders a way to compare code-generated video quality, and the voice API adds higher-fidelity identity control. The limitations are equally practical. The LiveAvatar demo is a starter, not a deployment; the voice-clone API has slot, account and recording-quality requirements; and the benchmark itself shows that following a brief is becoming easier than producing polished composition, timing and design craft.