Xiaofeiyu Translator E4B (小肥鱼 E4B 302ms 通用声带翻译核 — Universal Voice Sidecar for DeepSeek / Claude / GPT)
6–8 GB 小显存显卡 / 云端大脑外挂首选:给任意前沿大模型(DeepSeek / Claude / GPT / Qwen)外挂「真人短气泡五滑条声带」,代码块与工具调用
0%改动。基于google/gemma-4-E4B-it(仅占用~5.3 GiB显存,RTX 3060 8GB/RTX 4060 8GB/ 轻薄本均可常驻),配合@xiaofeiyu/openai-voice-proxy流式 AST 切分器,将上游大模型的说明散文在302 ms中位延迟内实时转译为小肥鱼五滑条短气泡口语,而把 Markdown 代码块、数学公式、URL/视频时间戳与tool_calls原封不动直通透传。
English Summary:
niatheelf777/xiaofeiyu-translator-e4bis a lightweight 4B (google/gemma-4-E4B-it, upper 20 non-QKV layersL22..41) 5-slider spoken-persona sidecar built for6–8 GBGPUs or users who want to keep DeepSeek, Claude, GPT, or Qwen as their upstream brain via@xiaofeiyu/openai-voice-proxy. Prose turns into natural multi-bubble Mandarin IM chat in~302 msmedian latency while fenced code blocks, identifiers, and tool calls pass through byte-for-byte untouched (0 / 36code-fence smuggling).
Xiaofeiyu (小肥鱼) Family Navigation / 家族导航
| 组件 / 仓库 | 定位与显存门槛 | 链接 |
|---|---|---|
| 12B 主力打磨版(端到端首选) | ≥ 16 GB 显卡 / 24 GB+ Mac 首选 Q8_0 原版挂 LoRA(~12.9 GiB);12 GB 显卡选特制 UD-Q4_K_XL-xfy(~8.5–8.9 GiB) |
xiaofeiyu-12b-v4(12B 主力打磨版) |
| E4B 通用声带翻译核 | 6–8 GB 显卡适用(~5.3 GiB),外挂 DeepSeek / Claude / GPT 零改代码以 302 ms 改写短气泡 |
xiaofeiyu-translator-e4b(当前仓 · E4B 通用声带翻译核) |
| 在线二维性格球试玩 | 免下载直接拖拽性格球体验五滑条实时混音与代码/算账分相 | https://xfy.tospeak.top |
| GitHub 一键本地启动仓 | 内置 :8081 纸墨风性格球网页 + OpenAI /v1 网关 + DSH 双模插件 |
github.com/niatheelf777/xiaofeiyu |
| HuggingFace 官方合集 | 小肥鱼主力底座与五槽口语卡带总览 | Xiaofeiyu Spoken Persona Collection |
Verified Benchmarks (Xiaofeiyu-Translator-E4B)
| Metric | Result | Details |
|---|---|---|
| Median 5-Slider Translation Latency | 302.1 ms |
129–140 tok/s on Q8_0 (~5.3 GiB total VRAM w/ 5 .gguf LoRAs) |
| Held-Out DPO Preference Accuracy | 96.05% (73/76) |
Zero-thought direct multi-bubble rewrite (Eval NLL 2.598 -> 1.368) |
| Code-Fence & Tool-Call Alteration | 0 / 36 (0%) |
Fenced code blocks, CLI flags, and JSON tool_calls bypass rewrite verbatim |
| Exact ID & Timestamp Preservation | 7 / 7 (100%) |
Video IDs, URLs, numbers, and timestamps preserved across all 5 sliders |
3-Minute Turnkey Quickstart (@xiaofeiyu/openai-voice-proxy + DSH)
git clone https://github.com/niatheelf777/xiaofeiyu.git && cd xiaofeiyu
pip install tokenizers huggingface_hub
# 1. Auto-download llama-server + gemma-4-E4B-it-Q8_0.gguf + 5 voice cartridges on 127.0.0.1:18085
python server/launch_local.py --preset e4b --llama-only --auto-download-llama
# 2. Launch zero-dependency OpenAI Voice Proxy on http://127.0.0.1:18088/v1
UPSTREAM_BASE_URL=https://api.deepseek.com/v1 \
UPSTREAM_API_KEY=your_key \
TRANSLATOR_URL=http://127.0.0.1:18085 \
node packages/openai-voice-proxy/bin/xiaofeiyu-voice-proxy.mjs
Point any OpenAI client, Cursor, or DeepSeek Harness (DSH) at http://127.0.0.1:18088/v1 (pass X-Xiaofeiyu-Sliders: 1.0,0.85,0,0,0 or install @xiaofeiyu/dsh-plugin for native multi-bubble rendering and the live 2D Persona Orb mixer):
node packages/dsh-plugin/scripts/install-dsh.mjs --apply --profile web --upstream-base http://127.0.0.1:18088/v1
5-Slot Voice Cartridges in This Repo
s0_xiaofeiyu_backbone.gguf— Slot 0 Everyday Warmth (元气日常)c1_neko_clingy.gguf— Slot 1 Sweet & Clingy (甜系软萌)c2_older_tease.gguf— Slot 2 Poised Older Sister (知性姐姐)c3_brat_mesugaki.gguf— Slot 3 Tsundere Wit (傲娇毒舌)c4_fragile_porcelain.gguf— Slot 4 Quiet & Gentle (安静慢热)
Base model usage is subject to the Gemma Terms of Use.
- Downloads last month
- 194
We're not able to determine the quantization variants.