Built in the open.

A second way to use your account: hold a key anywhere on Windows, speak, release, and the text lands at your cursor. Pairing is a one-time code you approve in the browser (no password ever enters the app), dictation shares your plan's existing allowance rather than adding a meter, and connected computers are listed on your account page where one click disconnects them. Installers for both regular PCs and Windows on ARM.

Recommend Cantari and keep 30% of every payment your referrals make in their first 12 months, on any plan and any billing period. The rates are published on the affiliate page rather than hidden, signup is instant with a real-time partner dashboard, and the rules (30-day window, last click wins, no self-referrals) are stated in plain language.

Both depend on a cloning engine we have not connected yet, so we are deferring them rather than shipping a feature that does not work. When they land, voice cloning will build a reusable voice from a short consented clip, and the voice changer will re-voice a recording as one of your clones with the timing and delivery intact. No invented launch date; we will say so here when it is real.

A forum beside the clip feed: ask for help and mark the answer that solved it, talk voices and engines, tell us what to build. Posts pass a safety layer that blocks only the severe stuff, holds the ambiguous middle for human review, and never touches honest criticism; everything we write carries the Team badge, including the first ten threads we seeded so you can see how it all works.

A public commons for things made here. Share a clip from your library or the studio (your work, your consent, engine and voice always credited), browse the feed, love what deserves it, and remix anything straight into your own studio with one click. A weekly prompt gives the commons a heartbeat; the first one is live.

Six hundred years ago a cantare was a story written to be performed aloud. That is the product, so now it is the name. The story lives on the about page, where it reads itself to you in Kore's voice while the text follows along.

Creator and Studio plans with flat monthly allowances, secure checkout, and a billing page that shows your real plan state. The free plan stays genuinely useful: drafting on Kokoro is unlimited on every plan, and that promise is now enforced in code, not just written on the pricing page.

Real product docs with a glossary, a supported-formats reference generated from the same configuration the studio runs on, per-format conversion guides, and a words-to-minutes calculator that can read your script aloud instead of just estimating it.

Every fixed demo clip on the site is now pre-generated once, verified word-by-word against its script, and served instantly: no waiting on a generation to hear a voice. New accounts get one question on arrival and land in a studio pre-loaded for what they are making, first clip one click away.

Docs, guides, engines, use cases, and posts now publish markdown twins, an llms.txt index, and a Copy page control with open-in links for popular AI assistants. Ask your assistant about Cantari and it can read the same honest pages you do.

Paste a whole manuscript or upload a text file: chapters are detected and split without altering your words, and an optional speakable pass expands numbers and abbreviations so engines read them right. Long books prepare in one pass with per-chapter status reported honestly.

The blog goes from coming soon to published: why we run an open benchmark, how we measure voice latency, and what owning your output means here. Every number in the posts is interpolated from the same data files the benchmark reads, so the writing cannot drift from the measurements.

Describe a mood and get an instrumental music bed or atmosphere from a text prompt, composed by Lyria 3 and saved straight to your library. The model is in preview, and one-shot sound effects are still coming.

Benchmark quality now uses the Artificial Analysis Speech Arena Quality Elo (a third-party, user-vote arena rating) instead of illustrative placeholders. Our Gemini engine rates within a few points of Fun-Realtime-TTS, the top model of all roughly 85 rated. Latency stays our own measured wall-clock number.

The roster grows from three engines to five. MAI Voice 2 brings the studio's first real fine controls (style, intensity, and speed parameters the engine genuinely accepts), and Zonos adds four American and British voices. Both carry a third-party Quality Elo on the benchmark (MAI matched to the nearest rated version), and latency is already measured.

A full studio editor with highlighted cues, Enhance Script, a searchable voice picker with live previews, and a real transport bar.

Record or upload your voice today; samples store privately and will be ready the day the cloning engine arrives.

Real transcription on a Whisper-class model, with file upload and direct in-browser recording.

Seven tool landing pages, a rebuilt home, benchmark and pricing depth, about and blog.

Sign in, save every generation privately, and track your monthly character allowance.

Real measured latencies published for every engine, with honest illustrative naturalness scoring.

The first version of the site, with a real multi-engine voice console: Gemini Flash, Kokoro, and Grok Voice generating from one studio.