Drop in AI-generated speech — an audio file, or a short video with an AI voice track — and Cadence gives it back with a human delivery shape: pauses that breathe, pacing that varies, the brittle top end softened, the level set for listening — and, for video, closed captions transcribed from the finished track. Same words. Same voice. Free, no account, no catch.
Audio: mp3, wav or m4a, up to 10 minutes and 25 MB. Video: mp4, mov or webm, up to 3 minutes and 100 MB — the picture comes back byte-for-byte untouched; only the voice track is processed. Video also comes back with closed captions, transcribed from the humanized track: embedded in the file for mp4 and mov, and downloadable as .srt and .vtt for every video. Five conversions a day per visitor, because free still runs on real computers.
The receipt is Cadence's signature: input and output hashes, the processing chain, and the before/after measurements — the claim, in a form anyone can check. Files are discarded 15 minutes after processing; download now.
| Measurement | Before | After |
|---|
Process-and-discard, by design: your file is processed and deleted — outputs are kept 15 minutes for download, then gone. Nothing is retained and nothing is trained on. Audio files are never transcribed; a video's voice track is transcribed on our own server only to make its captions, and that transcript is discarded with everything else. There is no account to attach anything to.
The Open Port mark beside the heading means: “An AI agent can connect here; tools are available at this point.”
If your agent speaks MCP, it can connect to Cadence directly — no upload form, no browser. Cadence runs as a local stdio MCP server, and any MCP-capable host can spawn it: Muse is one example, and the server is not tied to any single host. Two ways in:
The Cadence listing on Glama carries the server's details and the install route Glama provides for MCP hosts.
Free, MIT-licensed, and small: Python 3.10 or newer, standard library only, nothing else to install.
git clone https://github.com/CYAN-ECLIPSE-TECHNOLOGIES-LLC/cyan-eclipse-cadence.git
Then give your host this configuration — replace the two paths with your clone's location and a writable folder for Cadence's state:
{"mcpServers": {"cyan-eclipse-cadence": {"command": "python3", "args": ["/absolute/path/to/cyan-eclipse-cadence/server.py"], "env": {"CADENCE_MCP_STATE_DIR": "/absolute/path/to/writable/cadence-state"}}}}
Your host starts the server with python3 server.py — from there, the seven tools below are available to the agent.
Plainly: the Cadence MCP server does not make audio. It is non-executing — it inspects scripts, voice data, and render records, verifies receipt chains, and records render proposals. It does not synthesize speech, it does not humanize audio files, and it makes no network calls. To humanize a file, use the upload tool at the top of this page; the MCP connection is how an agent works with Cadence itself.
Cadence was built to make synthesized speech perform a script the way a person would — measured against real human narration, under rules cited from a century of elocution research. This tool applies the same discipline to finished audio:
Machine speech joins phrases with metronome gaps. Cadence finds the pauses and re-times them the way people actually pause — short breaths, longer emphasis pauses, a real stop at a boundary — so the delivery stops sounding typed.
People breathe before a phrase, and listeners hear it. A soft breath cue is laid into the longer pauses — the smallest sound in the chain, and the one that most often gives a synthetic voice away by its absence.
Each phrase gets a small tempo and level nudge, inside strict bounds, so the rate stops being machine-constant — without ever crushing the whole file faster or slower.
The glassy top end typical of AI renders is softened, the phrases are evened out, and the result is normalized to spoken-word listening level — ready to post, with the before/after numbers shown above.
Video conversions are transcribed after humanizing, on our own server — so the captions time what you actually hear. mp4 and mov files carry the captions inside the file, ready to toggle in any player; every video also gets .srt and .vtt caption files to download. WebM keeps its captions as those files, because players do not expose captions embedded in WebM.
It humanizes delivery. It does not re-perform your script, change your words, or turn one person's voice into another's — it has no ability to imitate a real person, by construction. Pauses are reshaped where they already exist. Video captions are transcribed automatically from the humanized audio, so an occasional word can be wrong — check them before you rely on them. Use it on your own AI-generated audio. On video, speech can shift by up to a third of a second as pauses are reshaped — it is made for voice-over video (reels, explainers, screen captures), not lip-sync dubbing.
One command gives your agent a catalog of MCP servers that carries its test results with it — what was verified, at what scope, and what was not tested. The same rule as this page: proof first, adjectives later.
This page is the front window. Cyan Eclipse Cadence — the performance compiler, measurer, and receipt engine behind it — is free and open source, and it connects directly to MCP-capable agents: the routes, the configuration, and the tool list are in the agent section above.
Produced with Cyan Eclipse Cadence — free at cetech.ai/free-tools