GPT-Live-1 vs HeyGen Voice：兩個看板、兩項任務，以及一個早已將兩者配對的 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.

每一個實際上是什麼

HeyGen Voice 是 HeyGen 自家研發的文字轉語音模型，內建於同時提供影片與虛擬人偶渲染的 v3 API 之中。目前發布了兩個模型 ID——heygen-voice-1，涵蓋即時克隆與專業克隆，以及heygen-voice-1-turbo，採用相同引擎但僅限用於即時克隆，延遲更低。在內部，這套引擎稱為orca，也就是你在 voices 查詢中用以篩選的值。它接收文字並回傳 44.1 kHz 單聲道 WAV 音訊，可以是一次性的請求（POST /v3/models/audio/tts），或是帶有 base64 WAV 片段及選用字元對齊事件的伺服器傳送事件串流（POST /v3/models/audio/tts/stream）。它不會聽見任何聲音。

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.

排行榜結果，以及其周圍的邊界

Artificial Analysis 經營兩個語音排行榜，它們回答不同的問題。Provider Voice 競技場讓每個廠商提交自己挑選的語音——公司挑選來代表自己的精緻示範語音——並根據這些語音聽起來有多好來排名。Controlled Voice 競技場則替換成一組固定的八個複製參考語音，四個美式、四個英式，因此每個引擎在相同條件下用相同語音說出相同的字詞。這是業界最接近同類引擎對等測試的東西，而且如果你打算複製一個語音並交給 TTS 模型，這個排行榜就是關鍵。

HeyGen Voice 領先。根據 2026 年 10 月的擷取：

• 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."

在更廣泛榜單上的缺席並不是品質訊號。Artificial Analysis 指出，條目可能因票數不足而暫不列入，而一個問世僅數週、背後只有一個評分方式不同的競技場的模型，尚未累積到九十多個模型的榜單所需的盲聽者投票量。正確的解讀既狹窄又有用：HeyGen Voice 目前是受控複製語音比較中得分最高的引擎，但相對於更廣泛的領域仍是未知數。任何人根據那個數字引用「世界上最好的 TTS 模型」，其實都是在引用一個比這句話所暗示更小的榜單。

HeyGen 為你提供什麼來打造

這個引擎比「TTS 模型」這個標籤所暗示的更靈活，而對任何要在這兩款產品之間做選擇的人來說，有三個細節很重要。

• 可為每個語音選擇引擎。GET /v3/voices 呼叫會接受一個引擎篩選條件，其值包括orca（代表 HeyGen Voice）、starfish，以及直接轉接至 ElevenLabs 語音的傳遞選項。HeyGen 不會強迫你採用它自家的模型——它以最字面上的方式分散風險，讓你透過它自己的 API 路由到競爭對手的引擎。任何把 1,198 Elo 解讀為「HeyGen 的語音最棒」的人，都應該注意：那只描述了同一個端點所提供的三條路徑其中之一。

• 複製功能提供兩種模式。Instant Clone 只需一段錄音，會使用前三分鐘，無法重新訓練，且文件記載在預覽期間免費。Professional Clone 會以一到十段錄音進行訓練，總時長二十多分鐘，會傳回 202 Accepted 並持續輪詢直到啟用，可重新訓練，且會佔用一個已購買的名額。相似度控制項的範圍為一到四，表情豐富度增強為 0.0 到 1.0，僅適用於 Instant 聲音，而專業層級則提供最長五秒的停頓標籤。

• 可重現的語音設計。POST /v3/voices會根據上限為 1,000 個字元的描述生成最多三個排序候選，並可選擇性加入性別篩選、BCP-47 語言地區設定與種子值——而 HeyGen 的文件說明，相同的提示加上相同的種子一律會回傳相同的語音。語音文字上限為 5,000 個字元，此端點在隨用隨付方案下每分鐘可執行 30 次請求，企業方案則為 120 次。

• 語言涵蓋範圍為 39 個基礎代碼，適用於複製與語音，地區標籤會解析至其基礎語言，而非切換口音。HeyGen 行銷頁面上流傳的平台層級「175+ 種語言和方言」數字，並非此模型的涵蓋範圍。

關於那份語言清單，有個實用警告：avatar 引擎所接受的幾個區域性 locale，在語音引擎中並不會選取區域口音。如果你的產品要進行本地化，請測試你實際需要的 locale，而不要假設該標籤會自動生效。

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.

當兩個系統交會之處：委派，以及其代價

雙方現在都提供完整的對話迴圈，而且都把推理器放在你能控制的位置。

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.

這週該如何決定

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.

不該做的是引用「Artificial Analysis 第一名」，彷彿這就已為某個對話式產品拍板定案。它解決的只是八個複製語音上的渲染問題。另一個問題在那個排行榜上根本還沒被提出。

這些都不是假設：像這次發布這類版本背後那些有紀錄的 AI 失效與代理事件，全都收錄在OrcaRouter AI Incident Archive，其中每筆條目都附有自己的來源、嚴重程度、信心等級與爭議標記。