User repository on GitHub
Correct VOICEVOX readings with TypeSafe Jev, and intonation as an option
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Integrations & Adapters
Voicevox Jev Proxy
A Python tool that asks TypeSafe Jev to fix Japanese readings in VOICEVOX: Sudachi splits sentences, Jev answers reading and phrase-boundary questions, and output comes as corrected WAV files or a VOICEVOX-compatible API server.
- Stars
- 2
- Language
- Python
- License
- MIT
- Reading time
- 2 min
- Published by
- Community
- Author
- nemalabs
- Last verified
- Oct 2026
Works forJev for Classification
- Languages
- Python
- Capabilities
- Open sourceRequires API key
Highlights
- Sudachi splits sentences and gathers reading candidates from UniDic and Sudachi dictionaries, with English senses from JMdict when present.
- One Jev request decides word boundaries, accent-phrase splits, readings, and whether www or 笑 signal laughter (笑 is read as written).
- Answers below a confidence threshold (0.6 default; 0.8 for phrase-height answers under --intonation) are discarded.
- Ships a VOICEVOX-compatible server on 127.0.0.1:50121: POST /audio_query returns corrected queries and everything else is forwarded.
- Needs Python 3.12+, uv, an engine at 127.0.0.1:50021, a TypeSafe key, and self-downloaded Sudachi and JMdict dictionaries; defaults to speaker 3.
Quickstart
bash
uv sync
uv run voicevox-jev-correct "雨が降ってきたから、傘を持っていくのだ。"
uv run voicevox-jev-proxyWatch out
Copy reviewed against the Japanese-language repo README on 2026-10-15. The README states intonation correction is still in development and off by default because accuracy is insufficient, and warns the proxy server has no authentication. Dictionaries are not bundled and must be downloaded separately.
Questions about this entry
Short answers drawn from the linked project or page.
It corrects Japanese readings before VOICEVOX speaks: Jev answers choice questions about where words split, which reading fits, where accent phrases break, and how to read ambiguous web slang, then the sentence is rewritten and re-read by VOICEVOX.
Sentences are split with Sudachi and reading candidates come from UniDic and Sudachi dictionaries, with English senses from JMdict. Jev answers the choice questions, and answers below the confidence threshold — 0.6 by default — are not used.
Python 3.12+ and uv, a VOICEVOX engine (default http://127.0.0.1:50021), a TypeSafe API key, and separately downloaded Sudachi system dictionary and optional JMdict_e.gz. Without the dictionaries it still runs but skips reading and phrase-splitting questions.
It also corrects intonation by asking a second set of questions about phrase attachment, sentence-final type and phrase relations. The README says this is still in development, so it is opt-in, and it doubles Jev requests per sentence; --requests 1 combines them.
Yes. uv run voicevox-jev-proxy listens on 127.0.0.1:50121; point a VOICEVOX-using app at that address and POST /audio_query answers with corrected queries while other requests pass through. It is not a full VOICEVOX replacement and has no authentication.
By default one Jev request per sentence, sent only when there is something to ask; --intonation makes it two. The server caps Jev requests per process with --request-cap (default 20, then 503), limits /audio_query text to 1000 characters (413 beyond), and allows 32 concurrent connections.
No. The README says there is no authentication, so binding the server to a non-loopback address with --host lets anyone reaching it spend your Jev requests and call every VOICEVOX API. Use --allow-host and --allow-origin for other machines and web pages.
More Jev resources like this
A session-aware, OpenAI-compatible gateway that routes chat completions across provider pools by cost, quality, and capability rules, optionally asks configured decision providers typed choice questions with a deterministic fallback, and records requests, decisions, and outcomes in SQLite.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
texasoct/jev-gatewayTypeScript
texasoct/jev-gatewayTypeScriptA Python library that runs LLM and Jev calls as Polars expressions: parallel inference across providers, Pydantic-validated outputs as struct columns, and a native TypeSafe System One layer for per-row decision questions.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
pnthn-ai/polar_llamaPython
pnthn-ai/polar_llamaPythonA GitHub App that routes pull-request review work through Jev: it ingests the webhook, normalizes the PR context, gets a LOW, NORMAL or SPECIALIST decision, and a deterministic policy layer maps it to auto-approval, a review request or a check-run update.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
hexyedev/jevprPython
hexyedev/jevprPythonCustom ComfyUI nodes that use Jev for text interpretation and judgments — select candidates, evaluate conditions, score text and extract numbers — plus a node for picking local SKILL.md files.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
hndrr/comfyui-jevPython
hndrr/comfyui-jevPythonA Claude Code mod that adds $.jev to the engine interface, letting plugin hooks ask Jev for Choice, Score and Noul answers over one state via TypeSafe, OpenRouter, Vercel, Cloudflare, LiteLLM or a custom endpoint.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
chrishan17/claude-jev-modTypeScript
chrishan17/claude-jev-modTypeScriptA host-agnostic routing layer that sits between a coding harness and its models, choosing capacity per round from trajectory evidence like tool calls, failures and context pressure, running as a user-scoped daemon and local OpenAI-compatible proxy.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
vizuh/sabiTypeScript
vizuh/sabiTypeScriptFrom 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.