Prosodia / index.html
lmcfarlin
Refresh demo: derisk-energy-24k milestone (energy direction, CFG amplification, long-form) replaces LJSpeech v1 samples
8a8f18e
Raw
History Blame Contribute Delete
7.53 kB
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Artificial Humanity — Prosodia</title>
<style>
:root {
--bg: #0f1115;
--panel: #171a21;
--ink: #e8e6e1;
--muted: #9a9790;
--accent: #d4a24e;
--accent-soft: rgba(212, 162, 78, 0.14);
--rule: #2a2e37;
}
* { margin: 0; padding: 0; box-sizing: border-box; }
body {
background: var(--bg);
color: var(--ink);
font-family: Georgia, 'Times New Roman', serif;
line-height: 1.65;
-webkit-font-smoothing: antialiased;
}
.wrap { max-width: 720px; margin: 0 auto; padding: 4rem 1.5rem 3rem; }
header { text-align: center; margin-bottom: 3.5rem; }
.masthead {
font-size: 0.78rem; letter-spacing: 0.28em; text-transform: uppercase;
color: var(--accent); margin-bottom: 1.2rem;
font-family: Helvetica, Arial, sans-serif;
}
h1 { font-size: 2.6rem; font-weight: normal; letter-spacing: 0.01em; margin-bottom: 0.8rem; }
.tagline { color: var(--muted); font-style: italic; font-size: 1.12rem; }
.rule { border: none; border-top: 1px solid var(--rule); margin: 2.6rem auto; width: 72px; }
h2 {
font-family: Helvetica, Arial, sans-serif; font-size: 0.8rem;
letter-spacing: 0.24em; text-transform: uppercase; color: var(--accent);
margin: 2.6rem 0 1rem; font-weight: 600;
}
p { margin-bottom: 1rem; }
.muted { color: var(--muted); }
.cast { display: grid; gap: 0.9rem; margin: 1.2rem 0; }
.role {
background: var(--panel); border: 1px solid var(--rule); border-radius: 10px;
padding: 1rem 1.2rem;
}
.role b { color: var(--accent); font-family: Helvetica, Arial, sans-serif; font-size: 0.95rem; }
.role p { margin: 0.3rem 0 0; font-size: 0.98rem; color: var(--ink); }
figure {
background: var(--panel); border: 1px solid var(--rule); border-radius: 10px;
padding: 1.1rem 1.2rem; margin: 1rem 0;
}
figcaption { font-size: 0.92rem; color: var(--muted); margin-bottom: 0.6rem; font-style: italic; }
audio { width: 100%; display: block; }
.note {
background: var(--accent-soft); border-radius: 8px; padding: 0.8rem 1rem;
font-size: 0.92rem; color: var(--muted); margin-top: 1rem;
}
.links { display: flex; flex-wrap: wrap; gap: 0.7rem; margin-top: 1.2rem; }
.links a {
font-family: Helvetica, Arial, sans-serif; font-size: 0.9rem;
color: var(--ink); text-decoration: none;
border: 1px solid var(--rule); border-radius: 999px; padding: 0.45rem 1.05rem;
transition: border-color 0.15s ease, color 0.15s ease;
}
.links a:hover { border-color: var(--accent); color: var(--accent); }
footer {
text-align: center; margin-top: 3.5rem; padding-top: 1.6rem;
border-top: 1px solid var(--rule);
font-size: 0.85rem; color: var(--muted);
font-family: Helvetica, Arial, sans-serif;
}
a { color: var(--accent); }
</style>
</head>
<body>
<div class="wrap">
<header>
<div class="masthead">Artificial Humanity presents</div>
<h1>Prosodia</h1>
<p class="tagline">The on-device dramatic audiobook engine.</p>
</header>
<p>
Most text-to-speech reads. Prosodia <em>performs</em>. It stages every book as a
production — read ahead, interpreted, and voiced with emotional direction — entirely
on your device. No cloud, no telemetry, no page of your book leaving your hands.
</p>
<h2>The Company</h2>
<div class="cast">
<div class="role">
<b>The Director</b>
<p>An on-device language model that reads ahead of the narration, annotating each
passage with emotional direction — valence, arousal, tension — the way a director
blocks a scene before the actors take it.</p>
</div>
<div class="role">
<b>The Actor</b>
<p>A neural voice trained by our <a href="https://huggingface.co/artificial-humanity/Sonora">Sonora</a>
project, performing the Director's notes through a Rust synthesis core — small enough
to live on a phone, expressive enough to be worth listening to.</p>
</div>
<div class="role">
<b>The Stage</b>
<p>A coordinator that keeps the performance flowing — gap-free audio, look-ahead
rendering, and graceful pacing across sentence and speaker boundaries.</p>
</div>
</div>
<hr class="rule" />
<h2>Hear the direction</h2>
<p class="muted">
Sonora’s de-risk milestone (July 2026): a multi-speaker, 24 kHz voice whose
<em>energy</em> can be directed on a continuous dial — the first trained channel of the
Director’s valence/arousal/tension vocabulary. Same text, same speaker, same seed; only
the direction changes. Artifacts in the
<a href="https://huggingface.co/artificial-humanity/Sonora">model registry</a>.
</p>
<figure>
<figcaption>Directed <b>hushed</b> (energy −1) — “The lighthouse keeper woke before dawn…”</figcaption>
<audio controls preload="none" src="samples/energy_minus1_spk245.wav"></audio>
</figure>
<figure>
<figcaption>Neutral (energy 0)</figcaption>
<audio controls preload="none" src="samples/energy_neutral_spk245.wav"></audio>
</figure>
<figure>
<figcaption>Directed <b>emphatic</b> (energy +1)</figcaption>
<audio controls preload="none" src="samples/energy_plus1_spk245.wav"></audio>
</figure>
<h2>Amplified direction</h2>
<p class="muted">
The Director can also push <em>past</em> the trained range with classifier-free
guidance — extrapolating the model’s own learned direction at render time.
</p>
<figure>
<figcaption>Neutral — “We won! We actually won the championship!”</figcaption>
<audio controls preload="none" src="samples/wewon_neutral.wav"></audio>
</figure>
<figure>
<figcaption>Energy +1</figcaption>
<audio controls preload="none" src="samples/wewon_energy_plus1.wav"></audio>
</figure>
<figure>
<figcaption>Energy +1, guidance ×3</figcaption>
<audio controls preload="none" src="samples/wewon_energy_plus1_cfg3.wav"></audio>
</figure>
<h2>Long-form stability</h2>
<p class="muted">
A single unbroken 103-second render — no drift in pace, loudness, or clarity.
Audiobooks need paragraphs, not sentences.
</p>
<figure>
<figcaption>~1¾ minutes of continuous narration, one pass</figcaption>
<audio controls preload="none" src="samples/longform_103s_spk245.wav"></audio>
</figure>
<div class="note">
Status: the energy channel is trained and verified (controllability ρ ≈ 1.0, speaker
identity preserved, intelligibility unchanged — the full evaluation ships with the
model). Valence and tension training is underway on the same plumbing, alongside a
247-voice casting space.
</div>
<h2>Follow along</h2>
<div class="links">
<a href="https://huggingface.co/artificial-humanity/Sonora">🎙️ Sonora model registry</a>
<a href="https://huggingface.co/artificial-humanity">🤗 Hugging Face org</a>
<a href="https://github.com/Artificial-Humanity">⌨️ GitHub org</a>
</div>
<footer>
Artificial Humanity — deeply expressive, local-first speech synthesis.
</footer>
</div>
</body>
</html>