Repository navigation
Releases
tsov v0.1.8 — the shape of voice
tsov v0.1.8 — the shape of voice
From v0.1.7 to v0.1.8 — the prototype learns a second language: the whole UI is now switchable between 中文 and English, every panel, canvas hint, dialog and status message included.
What's new since v0.1.7
- English UI (i18n). Switch 中文 / English under ⚙ Settings → 语言 / Language — or open the app with
?lang=en. Chinese stays the default and the choice persists across restarts. The entire UI shell is translated (roll, score, chat, docks, chain panel, canvas text, dialogs), with terminology aligned to the bilingual glossary (e.g. 试听 / audition, 转乐谱 / convert to score). User content — project names, chat messages, note data — stays untouched, as it should. - Docs — the README (both languages), glossary and feature list now document the language switch; upstream attribution refreshed (demucs, python-audio-separator, ADTOF, YourMT3).
- Fixes — two layout glitches surfaced by longer English labels: the chat header row no longer overlaps, and the input placeholder fits its row.
One-click bundle — a prebuilt Windows (win64) one-click package is distributed via the Quark Drive mirror; grab the current link from the README. (The bundle is larger than GitHub's release-asset limit, so it ships via the mirror.)
Under the hood — 746 tests, all green. The English build is gated by an automated check: a scripted scan of the rendered DOM must find zero Chinese characters outside content / user-data zones, with the Chinese build verified side by side. Every feature remains operable through the same transactional command layer.
Honest status: still a prototype. Windows-first; the UI ships in 中文 / English (switchable, Chinese default — no browser-language auto-detection). A few browser-native dialogs (confirm / prompt) still follow the system language. The transcription / rendering engines ship preconfigured inside the bundle.
License: AGPL-3.0.
tsov v0.1.7 — the shape of voice
tsov v0.1.7 — the shape of voice
From v0.1.5 to v0.1.7 — the prototype gets ears for finished songs and a wire to other agents. Two big lanes land: a reverse pipeline that pulls a finished recording apart and rebuilds it as an editable project, and an MCP interface that opens the studio to external agents.
What's new since v0.1.5
- Song decomposition — finished songs in. Drop a finished recording into the chain and hit run: it detects the tempo and aligns the grid, separates four stems (fast and high-quality modes), transcribes vocals with GAME, drums with ADTOF, and bass / everything-else with MT3 — then quantizes and snaps to key, and one click places the whole thing into the project as a multi-track arrangement. A 13-step chain, addressable step by step.
- MCP interface (experimental).
tsov mcpruns a stdio MCP server: external agents (Claude Code, Hermes, Codex, …) can list, read, edit, render and export projects, and fetch the built-in skills as prompts — through the same action layer the built-in agent uses. Edits stay reviewable and reversible; the write plane is off by default (--allow-write), with a project lock and a web-coexistence guard. - Bundle robustness (v0.1.6) — the one-click package ships ffprobe and degrades gracefully when git is absent.
- Fixes & polish — SSE self-healing reconnect (the root cause behind a "clicks do nothing" report), folder M/S indicator restored, chain-panel horizontal scroll now survives live re-renders, and the stem cache self-heals when files are briefly locked mid-render.
One-click bundle — a prebuilt Windows (win64) one-click package is distributed via the Quark Drive mirror; grab the current link from the README. (The bundle is larger than GitHub's release-asset limit, so it ships via the mirror.)
Under the hood — 746 tests, all green; every feature — UI control, REST endpoint, or agent tool — is verifiable through the same transactional command layer. The MCP surface was validated end-to-end by an external agent driving the full pipeline (humming → transcription → project → mp3).
Honest status: still a prototype. Windows-first; the UI is Chinese-first (English i18n planned). The transcription / rendering engines ship preconfigured inside the bundle.
License: AGPL-3.0.
tsov v0.1.5 — the shape of voice
tsov v0.1.5 — the shape of voice
From v0.1.0 to v0.1.5 — the workstation gets hands. Everything you touch and everything the agent touches moves through the same command layer — and now you can drag objects straight into the conversation.
What's new since v0.1.0
- Direct object references — drag a track, a selection, or an effect card into the chat input; it lands as a reference card and the agent receives a fully resolved object, with graceful fallbacks when things go stale. Favorites can now be renamed in place.
- Single-track workstation v2 — new gesture family (click keeps the main track · Ctrl-click switches-and-keeps · double-click switches away · double-click in the partner region swaps); the partner region hosts MIDI, audio, or lane curves; a lane system v2 with a unified lane renderer; and a folder layer (volume / mute / solo for folders).
- Lanes & editing — velocity lane, CC / pitch-bend lanes with proper MIDI export, effect parameter editor with registry-backed sliders and per-parameter reset, track & folder drag-reorder, triplet grids (8T / 16T), rubber-band move for selections, and waveform zoom that re-fetches peaks at each zoom level.
- Trust fixes — replace-range semantics, track addressing, velocity hints in prompts, level-summary wording, message double-display, branch-aware undo in the journal.
- Liveness — streaming heartbeat & progress states, thinking-strip UI, per-tool live adoption with a three-color diff overlay while the agent works, listen-while-arranging, and a real root-cause fix for the "frozen page" (O(n²) streaming layout).
- Hands-on fixes — drag-and-drop into chat (effectAllowed × dropEffect mismatch), horizontal wheel / middle-button panning in the roll view, folder fold now hides its lanes.
One-click bundle — a prebuilt Windows (win64) one-click package is distributed via the Quark Drive mirror; grab the current link from the README. (The bundle is larger than GitHub's release-asset limit, so it ships via the mirror.)
Under the hood — 686 tests, all green; every feature — UI control, REST endpoint, or agent tool — is verifiable through the same transactional command layer.
Honest status: still a prototype. Windows-first; the UI is Chinese-first (English i18n planned). The transcription / rendering engines ship preconfigured inside the bundle.
License: AGPL-3.0.
tsov v0.1.0 — the shape of voice
tsov v0.1.0 — the shape of voice
First public release. Hum it → get a song you can keep editing.
What's inside
- Humming quick-lane: denoise → loudnorm → transcribe → quantize → scale-snap, with artifacts kept at every step
- RMVPE (humming) + GAME (produced vocals) transcription backends
- Agent chat editing: action cards, diffs, one-click undo (bring your own LLM key — DeepSeek-compatible by default)
- Arrangement presets; mixing (buses / effect chains / automation / metering); export matrix (master / stems / buses / MIDI · 16-24-32 bit)
- Full browser workspace — FastAPI + zero-build vanilla ES modules · 600+ tests
Demo (assets attached to this release)
tsov-demo-v1-en.mp4— 2:14 demo, English subtitlestsov-demo-v1-zh.mp4— 2:14 中文演示(B站版另发)tsov-demo-song-amazing-grace.mp3— the demo song (Amazing Grace · folk arrangement · 320 kbps)
Every frame of the demo comes from the real workflow — no mockups.
Honest status: v0.1 is a prototype. Windows-first; the UI is Chinese-only (English i18n planned for v0.2); the transcription/rendering engines (FluidSynth / RMVPE / GAME) need manual setup — see the README.
License: AGPL-3.0.