dasheng

Checked, not fully reviewed. The speech half needs a self-hosted ASR model and about 12 GB of VRAM, which this machine does not have. We ran its tests and drove its decision module against a stand-in; we did not speak to it.

You read a sentence aloud, the words light up as the streaming recogniser catches them, wrong ones turn red as you go, and a score appears at the end — the pitch is an open replacement for a per-minute subscription app, with everything on your own card. Six tests pass (lib/align.test.js, the alignment between what you read and what was heard).

The Jev half is one module, and it is the tidy part: one request carries a question per word, so a whole sentence’s verdicts are a single round trip rather than one call per word. Driven against our stand-in, the request is exactly what you would hope:

body keys: ['state', 'model', 'questions']   state: "CANARY-SPOKEN-TEXT the quick brown fox"

The transcript text and the questions, nothing else. What leaves the machine is therefore what you said, and where it goes is worth reading twice: the default endpoint is not TypeSafe but a ZenMux gateway (zenmux.ai/api/v1/systemone), with JEV_ENDPOINT and JEV_API_KEY overriding it. A comment explains a nicety we appreciated — JEV_API_KEY is checked before ZENMUX_API_KEY precisely because another project’s gateway key may already be in your shell.

Chinese documentation. The proxy handling in the same module is a small piece of real-world engineering: Node’s global fetch ignores https_proxy, so it falls back to undici’s proxy agent with an explicit TLS servername, “because without SNI the local Clash drops the connection”.

For the same one-request-many-questions shape without a GPU, see Shapeshift; for another self-hosted audio boundary, call-coach-ai.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 27, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.