VoiceStudio評測：開源免費版ElevenLabs，聲音克隆全跑在本機 | VoiceStudio Review: Free Local ElevenLabs Alternative

9月 07, 2026

By Kit 小克 | AI Tool Observer | 2026-09-08

🇹🇼 VoiceStudio評測：開源免費版ElevenLabs，聲音克隆全跑在本機

VoiceStudio 這幾天在 GitHub 熱榜衝到全球前段班，短短一週多了超過 7000 顆星、累積衝破 1.9 萬顆星，主打「完全本地、開源免費」的 ElevenLabs 替代方案——聲音克隆、配音、口述聽寫、有聲書製作全部在自己電腦上跑，不用月費也不用把錄音上傳到別人的伺服器。

VoiceStudio 是什麼

開發者 debpalash 把 VoiceStudio 做成一個桌面應用，背後整合了 16 套文字轉語音引擎和 11 套語音辨識引擎，支援 646 種語言的字幕與配音。它還開了一個相容 OpenAI 格式的本地 API（localhost:3900/v1），讓 Claude Code、Cursor 這類工具可以直接呼叫語音合成或轉錄功能，對做內容自動化的人算是個實用彩蛋。

硬體門檻：GPU 不是必要，但會決定體驗

官方文件寫得算誠實：GPU 是選配，不是必需。純 CPU 也能跑，但速度會明顯變慢；想要接近即時的克隆速度，建議至少 8GB VRAM 的 NVIDIA 顯卡，部分大模型甚至要 12～16GB。硬碟則要預留約 10GB 給程式、Python 環境和模型權重。換句話說，這不是雙擊就能用的傻瓜軟體，比較像是給願意花半小時裝環境的工程師和內容創作者。

跟 ElevenLabs 比，誠實地說

克隆音色只需要 5～15 秒乾淨的人聲樣本，屬於零樣本（zero-shot）克隆——樣本更長不代表音質更好，錄音裡有背景音樂或殘響反而會拖累效果，這點跟付費雲端服務的體驗其實很接近。但老實講，VoiceStudio 在多語言的自然度和穩定度上，目前還是不如 ElevenLabs 這種打磨多年的商用服務；換來的是零月費、資料完全不出本機，而且內建 AudioSeal 浮水印技術可以標記合成語音，多少緩解了深偽疑慮。授權採用 AGPL-3.0，如果你拿它改機魔改後架設成網路服務對外提供，也得公開原始碼。

對於偶爾要幫 YouTube 影片配音、做 Podcast 轉錄、或單純不想把聲音資料交給第三方的人，VoiceStudio 值得抓來試試看；但如果你要的是穩定的商業級多語配音，現階段雲端服務可能還是比較省心。

好不好用，試了才知道

🇺🇸 VoiceStudio Review: Free Local ElevenLabs Alternative

VoiceStudio is the open-source, fully-local ElevenLabs alternative that has been climbing GitHub's trending charts this week, adding over 7,000 stars in a matter of days to cross 19,000 total. It packages voice cloning, dubbing, dictation, transcription, and audiobook creation into a desktop app that runs entirely on your own machine — no subscription, no uploading your voice to someone else's server.

What VoiceStudio Actually Does

Built by developer debpalash, VoiceStudio bundles 16 text-to-speech engines and 11 speech-recognition engines under one roof, covering 646 languages for captions and dubbing. It also exposes an OpenAI-compatible local API at localhost:3900/v1, so tools like Claude Code or Cursor can trigger voice synthesis or transcription directly — a genuinely useful feature if you are automating a content pipeline.

Hardware Requirements: GPU Is Optional, But It Matters

The documentation is refreshingly honest here: a GPU is optional, not required. CPU-only mode works but is noticeably slower; for near-real-time cloning, plan on an NVIDIA card with at least 8GB VRAM, with some larger models wanting 12-16GB. Budget around 10GB of disk space for the app, Python environment, and model weights. This is not a double-click-and-go tool — it is built for developers and creators willing to spend 30 minutes setting up their environment.

Honest Comparison With ElevenLabs

Voice cloning needs just 5-15 seconds of clean reference audio and works zero-shot — a longer sample does not guarantee better quality, and background noise or reverb will hurt results, which is roughly on par with paid cloud services. But to be blunt, VoiceStudio's multilingual naturalness and consistency still trail a polished commercial product like ElevenLabs that has had years to refine. What you get in exchange is zero monthly cost, data that never leaves your machine, and built-in AudioSeal watermarking to flag synthetic audio. It is licensed under AGPL-3.0, so if you modify it and run it as a network service for others, you are required to release your source too.

If you occasionally need to dub a YouTube video, transcribe a podcast, or just do not want your voice data sitting on someone else's server, VoiceStudio is worth downloading. But if you need production-grade, stable multilingual dubbing at scale, cloud services are probably still the lower-friction choice for now.

好不好用，試了才知道

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士