-
ASR: ONNX-exported Whisper-small (244 M params) → int8 quant → 30 % speed-up via TensorRT.
-
NMT: Transformer-base 12-layer, pruned to 320 M params; supports 200 language pairs, size 180 MB per pair.
-
TTS: VITS-based neural vocoder → 22 kHz natural prosody; latency 180 ms on Pixel 6 GPU.
Offline Strategy
Language packs are differential downloads—if you already have Spanish→English, adding Galician→English only pulls the 40 MB delta. All inference runs on-device; server hit happens only for obscure dialect crowdsourcing.
Language packs are differential downloads—if you already have Spanish→English, adding Galician→English only pulls the 40 MB delta. All inference runs on-device; server hit happens only for obscure dialect crowdsourcing.
Visual OCR Bonus
We splice EAST text-detection + TrOCR for handwritten menus; achieves 92 % acc on IAM-HW benchmark. AR overlay uses OpenGL ES 3.1 so it also works on four-year-old mid-range Samsungs—crucial for emerging markets.
We splice EAST text-detection + TrOCR for handwritten menus; achieves 92 % acc on IAM-HW benchmark. AR overlay uses OpenGL ES 3.1 so it also works on four-year-old mid-range Samsungs—crucial for emerging markets.
Monetisation Without Surveillance
No ads, no data resale. Revenue is pure freemium: offline packs free for 30 days, then $0.99 per language-year. Start translating → transgoing.com
No ads, no data resale. Revenue is pure freemium: offline packs free for 30 days, then $0.99 per language-year. Start translating → transgoing.com
ASR：ONNX 导出的 Whisper-small（2.44 亿参数）→ int8 量化 → 通过 TensorRT 提升 30% 的速度。
NMT：基于 Transformer 的 12 层模型，精简至 3.2 亿参数；支持 200 个语言对，每个语言对大小为 180 MB。
TTS：基于 VITS 的神经声码器 → 22 kHz 自然韵律；在 Pixel 6 GPU 上延迟为 180 毫秒。
离线策略
语言包采用差异化下载——如果您已经安装了西班牙语→英语，添加加利西亚语→英语只会下载 40 MB 的增量。所有推理都在设备端运行；仅在众包小语种时才会调用服务器。
视觉 OCR 附加功能
我们将 EAST 文本检测与 TrOCR 相结合，用于手写菜单识别；在 IAM-HW 基准测试中达到 92% 的准确率。 AR叠加层使用OpenGL ES 3.1，因此即使在四年前的三星中端机型上也能流畅运行——这对新兴市场至关重要。
无监控的盈利模式
无广告，无数据转售。收入完全来自免费增值模式：离线包免费试用30天，之后每种语言每年收费0.99美元。开始翻译 → transgoing.com
本文来自投稿，不代表独立开发前线立场，如若转载，请注明出处：https://91wink.com/transgoing-building-a-real-time-polyglot-that-works-in-airplane-mode-stack-snapshot/
