Speak first. Improve fast.先开口，再快速优化。
Describe what you see in English, then get transcript review, a more natural rewrite, and a stronger next answer you can reuse immediately in the next round of practice.
A small loop for everyday spoken expression一个面向日常表达的小循环
MollyTalk focuses on the moment after you speak: keep the transcript editable, explain what can improve, and give you a better next answer so practice does not stop at one imperfect attempt.
Browser speech recognition when available, with manual fallback when it is not.
Heuristic feedback layer for transcript, rewrite, improvement points, and a suggested next answer.
A lightweight practice page that works as a quick entry point before a deeper native experience.
0:00
If you pin this page to your home screen, the quick capture route becomes a low-friction speaking practice entry point.
把这个页面固定到主屏幕后，快速录音页就会变成一个低摩擦的口语练习入口。
Voice-first capture语音优先采集
Start talking naturally. If speech recognition works in your browser, the transcript will appear below in real time. If not, edit or paste your response manually and keep the demo moving.
Recommended practice move: describe a space, object, interface, or scene you can currently see, then use the response layer below to improve your next answer.
Minimal, honest, and recoverable最小、诚实、可修正
This version deliberately avoids fake accuracy claims. It focuses on a usable practice loop: speak, review, edit the transcript when recognition is wrong, and continue with a stronger next answer.
01
Speak naturally first instead of writing a perfect sentence before practice starts.
02
Recover quickly when speech recognition is imperfect by editing the transcript directly.
03
Use the rewrite and next-answer suggestion to practice one more round immediately.
--
Not analyzed / 尚未分析
Speak or paste a response first. / 请先开口或粘贴一段回答。
- Waiting for input... / 等待输入...
Waiting for transcript... / 等待转写...
Waiting for transcript... / 等待转写...
Not exam prep first, but everyday spoken expression.不是先做应试训练，而是先做好日常口语表达。
The original MollyTalk idea started from a simple frustration: users often know what they want to say, but not how to say it naturally. So the experience should start with speaking, then turn what they said into feedback they can immediately reuse in the next round.
Fastest path to a real link, not a perfect native rebuild.先拿到一个真实可访问链接，而不是先重做完美原生版。
This demo is intentionally web-first because it is the quickest way to produce a live, shareable artifact with a domain. It proves the core loop first, then leaves room to discuss how a deeper native experience could evolve from here.
A real spoken-English problem, shaped into a small practice tool一个真实口语痛点，被做成一个小练习工具
Scenario
The intended moment is simple: I point my phone at a scene, object, or interface, describe it in English, then immediately see what I expressed, what sounds weak, and how to answer better in the next round.
Real origin
The original oralPartner notes were not about exam prep. They were about a much more everyday frustration: I often know what I want to say, but I do not know how to say it naturally.
Current version
This web version keeps the core loop accessible: browser speech recognition when available, manual transcript editing, lightweight feedback, and a suggested next answer.
What it does not fake
MollyTalk does not claim production-grade pronunciation scoring. It gives lightweight feedback and keeps the user moving when transcription or phrasing is imperfect.
2023-12-25 STT devlog
Expo dev client, EAS local build, bundle identifier setup, fastlane dependency, and the config-plugin mismatch around@react-native-voice/voice.
2025-10-28 reflection
A later note confirms the core product judgment: browser transcription felt too weak for continuous voice practice, so the experience made more sense as iOS-first.
Original oralPartner concept
The archived note already contained the product direction: free talk first, correction second, partner not teacher, and feedback that pushes the next answer instead of ending at one-time grading.
The hardest part was reliability, not just API access.最难的不是接进来，而是让链路可靠。
In the native experiment I used Expo plus @react-native-voice/voice. The friction came from dev-client setup, local iOS build requirements, and a stale config-plugin dependency that needed an override.
Do not pretend recognition is perfect. Make recovery easy.不要假装识别完美，要把修正做得顺手。
My product answer to imperfect STT is not a fake confidence number. It is transcript preview, manual edit fallback, clearer phrasing guidance, and a next-answer suggestion so the user can keep practicing instead of getting blocked by one bad recognition result.
The useful unit is not raw transcript. It is reviewable speaking chunks.有用的单位不是原始文本，而是可回看的表达片段。
In this demo, chunking is intentionally lightweight: browser transcript segments plus sentence-level review. If I continued the product, I would combine pause boundaries, filler detection, and editable chunks to help users review one thought at a time.
- Show a believable voice input → feedback → next-turn product loop.
- Explain why continuity and recoverability matter more than cosmetic AI polish.
- Demonstrate that the original concept came from a long-term user pain, not a random feature idea.
- Move the strongest capture experience back to native iOS for continuous speaking sessions.
- Replace heuristic scoring with LLM-backed feedback plus clearer evidence for why a sentence feels unnatural.
- Add scenario packs such as describe-a-room, describe-an-interface, and respond-to-a-photo.