GPT-Live-1 与 HeyGen Voice：两块板，两项任务，以及一个已经将它们配对的 API

Most write-ups of HeyGen Voice open with the leaderboard result, and the result is real: first place on Artificial Analysis's Controlled Voice arena at 1,198 Elo, ahead of Q​wen-Audio-3.1-TTS-Plus at 1,186 and Eleven v4 Turbo at 1,162. G​PT-Live-1 is not on that board and never will be, because the board measures how a voice sounds and G​PT-Live-1 is the thing that decides when to stop talking. The more useful fact is that HeyGen's own engineers already know this: in September 2026 they shipped an open-source real-time avatar stack that uses G​PT-Live-1 as the voice and conversation layer and their own avatars around it. The two products are not competitors. They are two halves of a system that one of the two vendors has already assembled in public.

每一个到底是什么

HeyGen Voice 是 HeyGen 自研的文本转语音模型，位于 v3 API 中，该 API 还公开了视频与虚拟形象渲染能力。已发布两个模型 ID——heygen-voice-1，涵盖即时克隆和专业克隆，以及 heygen-voice-1-turbo，同一引擎但仅限即时克隆，延迟更低。在内部，该引擎被称为 orca，这是你在 voices 查询中用于筛选的值。它接收文本并返回 44.1 kHz 单声道 WAV 音频，既可以作为一次性请求（POST /v3/models/audio/tts），也可以作为服务器发送事件流，包含 base64 WAV 分片以及可选的字符对齐事件（POST /v3/models/audio/tts/stream）。它不具备听觉。

G​PT-Live-1 does nothing else. It handles a spoken conversation while a delegated backend model or agent reasons and calls tools, it listens while it speaks, and it runs on exactly one endpoint — v1/live/sessions — with Chat Completions, Responses, Realtime and Batch all unsupported. Audio and text in, audio and text out, images and video out of scope, knowledge cut-off 31 July 2025. Voice sessions bill at $0.05 per minute, by the second, with backend usage charged separately at normal model rates.

The architectural consequence is the whole comparison. On a HeyGen-style pipeline, the reasoner sits between the microphone and the voice. On G​PT-Live-1, the voice model is the pipeline and the reasoner hangs off it.

排行榜结果及其周围的边界

Artificial Analysis 运营着两个语音榜单，它们回答的是不同的问题。Provider Voice 竞技场允许每家厂商提交自己选定的语音——即公司挑选出来代表自己的精修演示语音——并根据这些语音的听感优劣进行排名。Controlled Voice 竞技场则改用一组固定的八个克隆参考语音，四个美式、四个英式，这样每个引擎都在相同条件下用相同的声音说出相同的内容。这是业界最接近同条件对等引擎测试的东西，如果你打算克隆一个声音并把它交给 TTS 模型，那么它才是真正重要的那个榜单。

HeyGen Voice 领先。来自 2026 年 10 月的截图：

• Controlled Voice Elo — HeyGen Voice 1,198 over 1,765 samples, against Q​wen-Audio-3.1-TTS-Plus at 1,186 over 2,097, Eleven v4 Turbo at 1,162, Eleven v4 at 1,159 and Cartesia Sonic 3.6 at 1,149.

• Provider Voice Elo — Eleven v4 Turbo leads at 1,330, Eleven v4 at 1,329, Q​wen-Audio-3.1-TTS-Plus at 1,296, Cartesia Sonic 3.6 at 1,285, Gemini 3.8 Flash TTS at 1,275. HeyGen Voice does not appear on this board.

• Price, on the arena's own basis — $30.00 per million characters, against Q​wen's $19.30 and Eleven v4 Turbo's $40.00. HeyGen publishes the same $30.00 per million characters on its own product page with the line "no packages, no minimums, no tiers."

在更广泛榜单上的缺席并不是质量信号。Artificial Analysis 指出，条目可能因投票不足而被隐去，而一个仅诞生数周、背后只有一个评分方式不同的竞技场的模型，尚未积累起九十多个模型的榜单所要求的盲听者数量。正确的解读既有限又有用：HeyGen Voice 目前是在受控的克隆语音对比中得分最高的引擎，但相对于更广泛的领域仍是一个未知量。任何根据那个数字引用“世界上最好的 TTS 模型”的人，引用的都是一个比这句话所暗示的更小的榜单。

HeyGen 为你提供了什么来构建

这个引擎比“TTS 模型”这一标签所暗示的更为灵活，而对任何要在这两款产品之间做选择的人来说，有三个细节至关重要。

• 每个音色可单独选择引擎。一次 GET /v3/voices 调用会接收一个引擎筛选参数，其取值包括orca（用于 HeyGen Voice）、starfish，以及直通转发到 ElevenLabs 音色。HeyGen 并不强迫你使用它自己的模型——它做了对冲，而且是最字面意义上的对冲：让你通过自己的 API 路由到竞争对手的引擎。任何把 1,198 Elo 解读为“HeyGen 的音色最佳”的人，都应注意：它描述的只是同一个端点所提供的三条路径之一。

• 克隆有两种模式。Instant Clone 接收一段录音，使用前 3 分钟，无法重新训练，且文档说明在预览期间免费。Professional Clone 基于 1 到 10 段录音进行训练，总计 20 分钟以上，返回 202 Accepted 并轮询直至激活，可重新训练，并占用一个已购买的槽位。相似度控制范围为 1 到 4，0.0 到 1.0 的表现力增强仅适用于即时语音，而最长 5 秒的停顿标签在专业层级可用。

• 可复现的语音设计。POST /v3/voices 可根据不超过 1,000 个字符的描述生成最多三个排序后的候选结果，并支持可选的性别筛选、可选的 BCP-47 语言区域和随机种子——HeyGen 的文档说明，相同的提示词加上相同的随机种子始终会返回相同的语音。语音文本上限为 5,000 个字符，该端点在按量付费方案下每分钟可处理 30 个请求，在企业版下为 120 个。

• 克隆和语音均覆盖 39 个基础代码，地区标签会归并到对应的基础代码，而非切换口音。HeyGen 营销页面上流传的平台级“175+ 种语言和方言”这一数字，并不代表本模型的覆盖范围。

关于那份语言列表，有一个实际提醒：虚拟形象引擎接受的若干区域语言环境，并不会在语音引擎中选择地区口音。如果你的产品要做本地化，请测试你需要的区域设置，而不是想当然地认为该标签会生效。

The part of G​PT-Live-1 that HeyGen's own stack demonstrates

On 16 October 2026 O​penAI released prompt-to-voice to all API customers: a POST to /v1/audio/voices with type: "prompt", a name, the description and "model": "auto". Creating a voice is free; using it bills at normal session rates. The response is metadata only — an ID, a name, a creation timestamp, your prompt — with no audio and no preview, so auditioning costs a session. The steerable qualities are accent, age, presentation, pitch, timbre, texture, loudness, speech rate and distinctiveness, all expressed in ordinary language with no numeric parameters, and O​penAI documents that generated voices are not guaranteed to be unique.

Put that next to HeyGen's design endpoint and the difference is not quality, it is engineering discipline. HeyGen gives you up to three candidates, a seed that makes results reproducible and a preview audio URL per voice. O​penAI gives you one voice, no preview and no uniqueness guarantee, under a 100-voice organisation cap that also applies to voices generated on behalf of your customers. If a voice is a build artefact you will need to regenerate, only one of these two surfaces is designed for that.

The preset libraries diverge just as sharply. G​PT-Live-1's Live voice guide lists twenty-one named presets plus marin as default; the nine added on 16 October are Nuri, Juni and Flitz in German, Brise and Sillage in French, Nira in Hindi, Harema and Shida in Japanese, and Noeul in Korean — which leaves Hindi and Korean with one voice each, both feminine. HeyGen's catalogue is documented at 300-plus pre-built voices across dozens of languages. Those are not comparable numbers in kind, and for a product that needs a wide cast of distinct speakers, the difference decides the decision on its own.

两个系统交汇之处：委托，及其代价

双方现在都提供完整的对话循环，并且都把推理器放在你可以控制的位置。

HeyGen's LiveAvatar ships in FULL mode, where HeyGen runs speech recognition, the language model, text-to-speech, turn-taking, memory and contexts, and LITE mode, where you keep your own agent and use HeyGen for the avatar and voice layer. G​PT-Live-1 does the same thing with a different boundary: Responses delegation, where O​penAI runs the backend model, or client delegation, where you point the session at your own harness, agent or service.

That shared shape is what makes the cost question identical on both sides, and it is the question most comparisons skip. On G​PT-Live-1, full active session time is billable including while the backend works. O​penAI's own cost guidance makes the arithmetic blunt: using a faster backend with parallel tool calls so the user finishes a minute earlier saves $0.05 in voice charges. The voice rate is a constant; session length is the dial, and on HeyGen's FULL mode the same economics apply to whatever reasoner sits in the loop. A TTS line item does not appear in a conversational product's budget in the way a procurement sheet implies — the reasoning minutes do.

That is the one place a routing layer changes the shape of the build rather than the price of a line item. Client delegation turns the backend into a variable you can route per conversation instead of a fixed dependency. OrcaRouter fronts 200-plus models at provider list price with no markup, which matters most exactly when a vendor cuts a price — the cut is live here the same day rather than at renewal. Automatic failover keeps a provider outage out of a live call, and the routing DSL lets you compose the reasoning step from several models rather than staking the session on one. To state the boundary plainly: OrcaRouter hosts neither G​PT-Live-1 nor HeyGen Voice, and neither appears in our catalogue. What it routes is the reasoning behind a delegated session.

本周如何决定

Write down which board your product is judged on, because these two models do not overlap there. If a human listens to the output asynchronously — a marketing video, a training module, an audiobook, a cloned brand voice reading many lines — you are on the Controlled Voice board, HeyGen Voice currently leads it, and its $30.00 per million characters sits mid-field against Q​wen's $19.30 and Eleven v4 Turbo's $40.00. Test it against both of those in the same arena, because two of the three are cheaper.

If a human talks back in real time, the Controlled Voice board has nothing to say to you. The question becomes who owns the turn-taking loop, what the backend costs per session, and whether the voice model is the pipeline or a component inside one. HeyGen's answer is LiveAvatar with an agent you bring; O​penAI's answer is a session that delegates. Both are viable, and the fact that HeyGen's own reference stack uses G​PT-Live-1 for the conversational half is the clearest evidence available that these are not substitutes.

不该做的，是引用“在 Artificial Analysis 上排名第一”，好像这就为某个对话产品的决定一锤定音。它定下的只是八个克隆音色上的一个渲染问题。另一个问题在那个榜单上压根就没被问过。

这一切都并非假设：像本次这样的发布背后那些有据可查的 AI 故障与智能体事件，均收录于 OrcaRouter AI 事件档案库中，其中每条记录都附有自己的来源、严重程度、置信度评级和存疑标记。