User repository on GitHub
大声读 — R2T2 流式 ASR 听，Jev 逐词判，英文朗读评分
- Forks
- 20
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
大声读 (ReadAloud)
An English reading tutor for Chinese learners: R2T2 streaming ASR transcribes as you read aloud with a 160 ms step and committed text that only grows, then Jev answers closed questions per mismatched word — same word or not, error category, sentence score.
- Stars
- 144
- Language
- JavaScript
- Reading time
- 2 min
- Published by
- Community
- Author
- wquguru
- Last verified
- Oct 2026
- Languages
- JavaScript
- Capabilities
- Self-hostedRuns locallyRequires API key
Highlights
- Read English aloud: words light up as they are recognized, misread words turn red on the spot, and a total score button appears at the end.
- R2T2 (Confucius4-R2T2, about 2B parameters) is truly streaming with a 160 ms minimum step and a stable prefix that only grows, so verdicts never flip.
- Jev answers only closed questions — same word or not with a probability, error category, and a 1-5 sentence score — and only mismatched words are sent.
- Everything runs self-hosted: a 12 GB NVIDIA GPU serves R2T2, audio stays on the machine, and Jev is reached through ZenMux's System One endpoint.
- ASR_PROVIDER=mock replays a scripted stream with three planted error types, so the whole scoring chain runs without a GPU.
Quickstart
bash
npm install
npm run dev
npm testWatch out
Copy reviewed against the repo README on 2026-10-15; the README is in Chinese and GitHub reports no licence. Jev judges text only: pronunciation, stress and accent are explicitly not scored, and R2T2 self-hosting requires CUDA because its infer_mode is tied to vLLM.
Questions about this entry
Short answers drawn from the linked project or page.
An open-source English reading tutor for Chinese learners. You read a passage aloud; streaming speech recognition transcribes it, words light up as recognized, misread words are marked red live, and a total score is produced at the end.
Jev handles the judgment step: per mismatched word it answers a noul question (same word or not) with a probability, a choice question for the error category such as substitution or self-correction, and a 1-5 score for whether the same sentence is still being read.
An alignment step in lib/align.js selects the words that do not line up; only those are worth asking a model about. Missed words, extra words, completeness, speaking rate and pauses are all arithmetic done locally.
For real ASR, self-hosting R2T2 needs an NVIDIA GPU with about 12 GB (Ampere or newer) and CUDA, because its infer_mode is tied to vLLM; Macs cannot run it. Set ASR_PROVIDER=mock to run the whole chain without a GPU.
Yes for scoring: you supply a ZenMux key as JEV_API_KEY and target the System One endpoint at zenmux.ai/api/v1/systemone with model typesafe/jev-1.13. Jev goes through its own System One interface, not chat/completions.
No. The README states Jev only reads text and cannot judge acoustics, so pronunciation, stress and accent produce no score; scoring those would require a separate phoneme-level pronunciation assessment model.
npm install, then npm run dev (port 3000, auto-switches if busy). npm test runs the offline alignment, mock and pacing unit tests without network. .env.local holds the JEV_* and R2T2_* settings and should not be committed.
More Jev resources like this
A localhost teaching aid that uses the real Jev API to review English instructional passages and estimate how many ideas a learner must hold at once, comparing the demand against a 1-5 slot budget with average, peak, and confidence labels.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
austinaway/working-memory-jevPython
austinaway/working-memory-jevPythonA local app that reads a few lines of notes about your product and lets Jev pick its way to a finished landing page: Jev never writes text, it only chooses between options, and code builds the page from those picks.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
sagapranav/jev-pagesHTML
sagapranav/jev-pagesHTMLA zero-label dining decision web app built on Jev's System 1 probabilistic decisions: type a plain-language craving like I got soaked and want hot soup, and it answers with a decisive pick, a two-way duel, a reasoned impasse, or a blind box.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
haoxiangxia/fork-castJavaScript
haoxiangxia/fork-castJavaScriptA read-only triage dashboard that finds likely duplicate GitHub issues: lexical retrieval shortlists candidates, Jev answers four Choice questions per pair, and selected passages are shown side by side. It never closes issues.
Jev Tools & CLIs#community#catalog-import#logicrw
An enterprise admin-console test tool that generates replayable Excel test cases from live browser pages, or replays existing Excel cases, driving Playwright with local Laya or a compatible decision API to match fields, options, and targets.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
anthonyli/laya-pilotJavaScript
anthonyli/laya-pilotJavaScriptWrite end-to-end tests for web apps in plain English; jev-e2e turns each case into a plan, has Jev pick the page controls, and runs it with Playwright, returning PASS, FAIL or BLOCKED with HTML, JSON and masked screenshot evidence.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
perixtar/jev-e2eTypeScript
perixtar/jev-e2eTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.