Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
| # Roadmap | |
| Goal: a public app that translates speech in any video, link or book — all languages, dialects and eventually ancient languages — on our own models. | |
| ## v0.1 (now) | |
| - [x] Link or upload → auto-detect → translate → transcript, subtitles, voice-over | |
| - [x] 108 target languages; offline NLLB model by default | |
| - [x] Optional Claude (quality) and ElevenLabs (realistic voice) | |
| - [x] "Request a language" form (saved to `data/language_requests.jsonl`) | |
| ## Next | |
| - [ ] **Speaker detection** — separate voices per speaker (male/female) automatically | |
| - [ ] **Voice that matches the speaker's tone** — only with the speaker's consent | |
| - [ ] **Lip-timed dubbing** — stretch/compress lines to mouth movement | |
| - [ ] **Live mode** — translate a stream or call in near real time | |
| - [ ] **Browser extension** — translate any video on the page you're watching | |
| - [ ] **Books & photos of pages** — OCR → translate → read aloud | |
| ## Own models / dialects (Lao & Isan first) | |
| - [ ] Collect Lao and Isan audio + text (with permission) → fine-tune Whisper so Lao isn't confused with Thai | |
| - [ ] Fine-tune NLLB on Lao↔English pairs from textbooks | |
| - [ ] Community review: native-speaker voting + paid reviewers + automatic quality checks | |
| - [ ] Turn language requests into a data-collection queue (Hawaiian, Native American languages, etc.) | |
| ## Going public | |
| - [ ] Accounts, usage limits, payments | |
| - [ ] GPU server (one GPU ≈ 20× faster than a laptop) | |
| - [ ] Storage cleanup, abuse/copyright policy, consent rules for voice | |