A Reader for Your Audiobooks, Two New Models, and Agent Voice
A word-synced Audiobook Reader, two new voice models, voice output for AI agents, correct math and number reading, and Qwen3 now runs about 4x faster. On desktop and in the cloud.
Here's what's new in Voice Creator Pro v2.0. This is our biggest release yet, so let's get into it. Everything below is live on both the desktop app and in VCP Cloud.
The Audiobook Reader
The headline feature. Audiobook projects now have a dedicated Reader, a reading view that follows the narration word by word.
- Word-synced highlight lights up each word as it's spoken, with smooth autoscroll that keeps the current line in view
- Click any word to select it, then play from that exact spot or copy the text
- Chapter rail for jumping around the book, plus a scrolling title page with the book title and author
- Full transport with a live seek bar you can scrub while streaming, a playback-speed control, and a sleep timer that only counts down while audio is playing
- A first-run wizard and settings for reading width, reading mode, and Light, Dark, or System themes that match the app
Open it from any project card or the editor header and read along while it narrates.
Two New Models
The lineup grew by two.
- VoxCPM2 is a high-quality 48 kHz model with cloning, voice design, and streaming. It comes with a seed control and a set of extreme-emotion starter prompts for expressive reads.
- Gepard 1.0 is a fast voice-cloning model with streaming support, great when you want a clone quickly.
Qwen3 Is About 4x Faster
Qwen3, the workhorse model, now runs through a new CUDA-graph fast decoder that makes generation roughly 4x faster, with time to first audio under half a second. Design mode uses it too, and audiobook projects render whole books in fast batches instead of one line at a time. Because it got so much faster, Qwen3 pricing in the cloud dropped to match.
Numbers and Math That Read Correctly
Spoken text now handles technical content the way you'd actually say it out loud. That means currency, dates, times, ordinals, fractions, percentages, phone numbers, scientific notation, and units, plus LaTeX, unicode, and plain-text math, operators, and Greek letters. Full number-to-words works across many languages, not just English. Paste in a research paper or a recipe and it just reads right.
Agent Voice
You can now give an AI agent a voice. VCP includes a built-in MCP server and an OpenAI-compatible speech endpoint (/v1/audio/speech), so an assistant or any tool that already talks to a speech API can speak in any of your voices. It's protected by an API key on every request and works safely over a tunnel, so your agent can reach it from anywhere.
Better Video Dubbing
Dubbing got a lot of attention this release.
- Per-speaker voice cloning clones each speaker from the source audio and re-voices them in their own voice, even splitting a label automatically when it turns out to cover two different people
- A preview toggle switches between Original, Dubbed, and Lip-synced while keeping your place in the video
- A translation glossary lets you pin how specific names and terms should translate, applied consistently across both dubs and subtitles
- A reworked editor and mixing pass with faster generation under the hood
Speaker Labels in Transcription
Speech-to-Text can now tell speakers apart. Multi-speaker transcripts label who's talking with colored badges, and the speaker labels carry through to your SRT and JSON exports.
A Redesigned Model Picker
Choosing a model is clearer now. A new info panel shows quality, speed, and price at a glance, along with an estimate of how much audio time you have left on each one, with languages ordered by how commonly they're used.
In the Cloud
Running VCP as a worker to earn money on your idle GPU is now hands-off. Workers update themselves automatically in the background and finish any job in progress before applying an update, so your node stays current without you touching it. The marketplace is still growing carefully, so reach out if you'd like in.