IdeaWell

Project ideas from Hacker News discussions.

Gemini-3.5-Transcribe

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Theme 1: Accuracy & hallucination issues
Users frequently complain about erratic sentence breaks, word removal, and hallucinations on silent or noisy input.
- jeffbee: “My biggest complaint is that it inserts sentence breaks in random places, then fails to evaluate the result, even though it is obviously wrong.”
- film42: “If you pass chirp some audio with noise or even no audio, it will barf text at you like ‘I don’t know. I don’t know. I don’t know.’ until a request timeout fires after like 10 minutes.”
- wolvoleo: “I've seen Whisper do this exact thing as well … repeating a few words over and over near the end … It’s really annoying.”
- Crystalin: “The main issue is when you want to say something precise … it might ‘simplify’ it and break the meaning. … ‘I hesitated to check it, I should have verified’ ⇒ ‘I should have verified’ (The ‘I hesitated…’ is removed).”

Theme 2: Latency is critical for real‑time use
Many commenters stress that low latency matters most for dictation, live transcription, and meeting notes, praising the model when it’s fast but noting it still lags behind competitors.
- k9294: “I am using it all day as a main dictation model, and I can say it's the best model in terms of accuracy, latency, and formatting I have ever used.”
- lnalx: “From language detection and accuracy in a noisy environment to the most important point: latency … Google released Gemini‑3.5‑Transcribe, and it beats every other model on accuracy, but it definitely needs a bit more work on latency, which is the most important factor for STT apps.”
- decimalenough: “Depends on your use case. If you're transcribing meeting notes, latency is a non‑issue.” (shows the use‑case split.)

Theme 3: Limited device availability & staggered rollout
The feature is presently tied to specific Pixel/GBoard releases and relies on cloud processing, leaving many users waiting or unable to access it.
- gumby271: “It says it's available in GBoard on Android but I don't see it yet. I assume like most Google features it will be a rolling release over the course of several months…”
- bahmboo: “As far as I can tell 'rambler for Android' is a named subfeature of gboard and only on Pixel 10+ and newer Samsung phones. From a user standpoint it 'just' seems like the transcription has gotten better …”
- blissofbeing: “I've been using Wispr Flow on my pixel 9 … it's available on my Pixel 9 unlike rambler that's currently only available for pixel 11.”
- hypfer / k9294: “Where does the compute happen? I suppose it's a cloud thing?” … “Yep.”

🚀 Project Ideas

TranscriptPolish

Summary

  • A desktop/web service that takes raw speech‑to‑text output (Whisper, Gemini, etc.) and runs a second‑pass LLM to fix hallucinations, restore missing content, correct random sentence breaks, and apply user‑defined style guides.
  • Core value: delivers clean, publication‑ready transcripts with minimal manual editing, saving time for journalists, researchers, and language learners.

Details

Key Value
Target Audience Content creators, journalists, researchers, language learners who rely on STT for notes or subtitles
Core Feature LLM‑based cleanup pass that detects and corrects hallucinations, restores truncated sentences, inserts proper punctuation, and applies style rules (e.g., capitalization, speaker labels)
Tech Stack Python backend (FastAPI), LLM inference via HuggingFace Transformers or Gemini API, optional Whisper.cpp for local STT, Electron or Tauri frontend for drag‑and‑drop UI
Difficulty Medium
Monetization Revenue-ready: Subscription $9/mo for unlimited cleanup, free tier limited to 1 hour/day

Notes

  • Addresses jeffbee’s complaint about random sentence breaks and hard‑to‑fix errors, and verdverm’s suggestion to run a second model to clean up transcripts.
  • Provides confidence scores and editable UI so users can flag systematic mis‑predictions (as totetsu wanted) and feed corrections back for personalized improvement.

DictationHub

Summary

  • A cross‑platform dictation app that lets users switch between local high‑quality STT models (Voxtral, Parakeet, Whisper) and cloud APIs (Gemini, ElevenLabs) with real‑time confidence scoring, inline correction, and adaptive language detection.
  • Core value: gives power users a fast, private dictation experience that never “simplifies” their speech and can be tuned for latency vs accuracy.

Details

Key Value
Target Audience Professionals, developers, and multilingual users who need reliable voice input on desktop/macOS (crossroadsguy, Crystalin, Lucasoato)
Core Feature Real‑time transcription with per‑word confidence highlighting, one‑click correction, custom glossary for industry jargon, and seamless fallback to cloud when higher accuracy is needed
Tech Stack Tauri (Rust) for low‑overhead UI, Whisper.cpp / ggml for local models, WebSocket client for Gemini/ElevenLabs APIs, React for settings panel
Difficulty Medium
Monetization Revenue-ready: One‑time purchase $29 (includes free cloud credits $5/mo)

Notes

  • Solves crossroadsguy’s search for a simple fast dictation app and Crystalin’s frustration with models that drop words (“I hesitated…”) by preserving original wording and offering verbatim mode.
  • Meets Lucasoato’s need for multilingual, industry‑term handling and totetsu’s request for confidence scores and correction loops, while providing an easy way to pay for cloud tokens without navigating Google Cloud Console (zhivota’s pain).

LiveSubtitleAI

Summary

  • An overlay utility that captures system audio or video files, runs streaming STT (choice of model), translates the transcript into the user’s language, and displays synchronized subtitles in real time.
  • Core value: enables language learners and non‑native speakers to enjoy any media with instant, accurate subtitles, while offering options to reduce hallucinations and adjust latency/accuracy trade‑offs.

Details

Key Value
Target Audience Language learners, expatriates, and anyone wanting live translated subtitles for movies, meetings, or YouTube (Freedom2, adamgoodapp)
Core Feature Streaming STT + translation pipeline with selectable models (Whisper, Gemini‑3.5‑Transcribe, Soniox), language detection, speaker diarization optional, and on‑screen text extraction via OCR for style‑guided subtitles
Tech Stack Electron or Tauri frontend, WebAssembly Whisper.cpp for local streaming, optional gRPC to cloud STT/translation services, React overlay, FFmpeg for audio capture
Difficulty High
Monetization Revenue-ready: Usage‑based pricing $0.006 per minute of audio processed (matches Gemini‑3.5‑Transcribe Live pricing) with free tier 30 min/day

Notes

  • Directly addresses Freedom2’s desire to check understanding with film/video via subtitles and adamgoodapp’s request for live translated subtitles for his Japanese wife.
  • Tack

Read Later