PuriPuly：VRChat 实时双向语音翻译

Real-time two-way voice translator for VRChat — LLM-based, VR subtitle overlay & OSC

PuriPuly Real-Time Two-Way Voice Translator for VRChat 🇺🇸 English · 🇰🇷 한국어 · 🇯🇵 日本語 · 🇨🇳 简体中文 · 🇷🇺 Русский Demo pph_clip_1_github.mp4 If you want to see more of actual communication with other foreign friends through PuriPuly: Demo 1 Demo 2 Demo 3 Finally, talk like real friends. You've been there. Wanting to comfort a friend, but only managing: "Are you okay?" You already know a 'translator' can't carry what's truly in your heart. So I built one that can. What is PuriPuly? PuriPuly is a two-way voice translator for Windows that translates your voice and the other person's voice in real time. We aim for natural translation through LLMs. Beyond stiff literal translation, so real person-to-person conversation can happen. It works in many environments, including VRChat and Discord. LLM-Powered Localization — Slang, colloquialisms, and casual/formal speech, all rendered naturally. Context Memory — Keeps the conversation flowing naturally with awareness of prior context. Two-way Voice Translation — Translates the other person's voice too, with VR subtitle overlay support. Start via Discord — Get going right away without a complex setup process. The most powerful local full stack — From Parakeet to Gemma 4 E4B, built with the most efficient models available today. Q&A How good is the translation quality? → You can comfortably have even the deepest conversations people can share. It also outperforms traditional commercial translation services by a wide margin. See the 'Translation Comparison' section below for details. How long does it take from speaking to getting a translation? → Under optimal conditions, latency is around 1 second, measured from when the speaker finishes talking. Does it cost money to use? → Yes, but only later. New users get a free usage allowance, and even after that the pricing is very cheap; you can translate thousands of times for $1. You can also use it for free by running local models. Do I need to get an API key? → Yes, but again, only later. At first, just install and authenticate via Discord to start using it. Voice recognition is slow. → When using local ASR, processing times can be slow if compute resources are limited. In this case, we recommend switching to a cloud STT service. How is personal data handled? → Voice and conversation contents are never sent to Puripuly servers. In addition, all source code is publicly available in this repository, so you can verify its network behavior directly. 📥 Download Translation Comparison The blue bars are the models available in PuriPuly. We ran the experiment using Microsoft's Gemba MQM framework. It was set up as a multi-turn environment to better resemble real conversation. For the full results, see here. Cost Uses per Dollar Recommended Models LLM \ ASR Local ASR Cloud Free Tier ASR Soniox Qwen Audio Index-Translate 35B A3B Unlimited Unlimited 5,000 7,260 Gemma 4 26B A4B + 31B 13,940 13,940 3,680 4,770 DeepSeek V4.1 Flash 16,800 16,800 3,860 5,070 GPT-6 luna 8,680 8,680 3,170 3,950 Other Models LLM \ ASR Local ASR Cloud Free Tier ASR Soniox Qwen Audio Gemma 4 E4B & 26B (Local) Unlimited Unlimited 5,000 7,260 DeepSeek V4 Flash (OpenRouter) 17,020 17,020 3,860 5,090 Gemini 3.8 Flash 1,160 1,160 940 1,000 Qwen 3.8 Flash 7,460 7,460 2,990 3,680 Cost per Utterance Recommended Models LLM \ ASR Local ASR Cloud Free Tier ASR Soniox Qwen Audio Index-Translate 35B A3B $0 $0 ~$0.0002 ~$0.00014 Gemma 4 26B A4B + 31B ~$0.00007 ~$0.00007 ~$0.0003 ~$0.00021 DeepSeek V4.1 Flash ~$0.00006 ~$0.00006 ~$0.0003 ~$0.00020 GPT-6 luna ~$0.00012 ~$0.00012 ~$0.0003 ~$0.00025 Other Models LLM \ ASR Local ASR Cloud Free Tier ASR Soniox Qwen Audio Gemma 4 E4B & 26B (Local) $0 $0 ~$0.0002 ~$0.00014 DeepSeek V4 Flash (OpenRouter) ~$0.00006 ~$0.00006 ~$0.0003 ~$0.00020 Gemini 3.8 Flash ~$0.0009 ~$0.0009 ~$0.0011 ~$0.0010 Qwen 3.8 Flash ~$0.0001 ~$0.0001 ~$0.0003 ~$0.00027 Based on (Input 900 tokens + Output 12 tokens) × 1.2 avg LLM calls per utterance. Uses per Dollar is derived from the un-rounded values in the Cost per Utterance table. All costs and usage counts are approximate. DeepSeek V4.1 Flash assumes a 70% cache hit rate; V4 Flash (OpenRouter) assumes 60%. GPT-6 luna uses OpenAI API pricing with no cache discount assumed; when connected through ChatGPT, it uses your Codex limits instead of API charges. Qwen API costs are based on the Beijing region. Pricing as of October 17, 2026. Free Credits Service Free Credit Duration Note Deepgram $200 None No card required ElevenLabs 10,000 credits Monthly reset No card required Gemini 3.5 Transcribe Free tier None Practically unlimited on the free tier Alibaba Cloud 1M tokens per model 90 days Singapore region Alibaba Cloud ¥300 1 year Students in China Local Models PuriPuly comes with the following local models built in. You can also connect OpenAI-compatible APIs. GPU inference runs on Vulkan. It works regardless of the vendor — Radeon or Arc alike. ASR Model Runtime Quantization Parakeet TDT 0.6B v3 CPU INT8 Parakeet TDT-CTC 0.6B (ja) CPU INT8 Qwen3-ASR 0.6B CPU INT8 Qwen3-ASR 1.7B GPU Q6_K OpenAI-compatible API — — LLM Model Runtime Quantization Gemma 4 E4B IT QAT CPU / GPU UD Q4_K_XL Gemma 4 26B A4B IT QAT GPU only (Vulkan) UD Q4_K_XL OpenAI-compatible API — — Managed Gemma weights download only when the selected model is needed; they are not bundled with the installer. The 26B download is approximately 14.25 GB. Model selection persists across restarts. E4B supports CPU with MTP or GPU; 26B requires a working Vulkan GPU and does not silently switch to CPU or E4B after a failure. Select another model/backend explicitly if the reported device or resource problem cannot be resolved. Self and Peer can translate concurrently on one managed GPU runtime, with at most two active requests. CPU translation remains serialized. The three retained language-pair prefix slots are separate from this active-request limit. See the managed local translation contract for model identities and fixed profiles. If you