Monorepo for Aesthetic.Computer
aesthetic.computer
say: add Jeffrey PVC voice provider + browsable utterance cache
New provider: `say:jeffrey hello` hits the backend with Jeffrey's
Professional Voice Clone (ElevenLabs voice dYNGZ848Oo6DtNBoeqgh, same
one used in the LACMA 2026 grant video). `say:jeffrey:scream` routes
to the screaming voice settings.
Cache layout changes (all still under art.aesthetic.computer/tts-cache/):
- Jeffrey utterances land in their own subfolder
`tts-cache/jeffrey/<sha256>.mp3` so they're easy to list and audit.
- Every cached object now carries the original text + provider + voice
+ timestamp as S3 user metadata, so browsing a HeadObject response
tells you exactly what was said and when. Values are ASCII-trimmed
and newline-stripped to satisfy S3 header rules.
Frontend (say.mjs):
- `jeffrey` added as a colon option (alongside google/openai/eleven)
- Provider indicator now renders in magenta when jeffrey is active
- Help line updated; README call-out documents the cache structure
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
+91
-11
+69
-5
system/netlify/functions/say.js
+69
-5
system/netlify/functions/say.js
···
29
29
const CACHE_PREFIX = "tts-cache/";
30
30
31
31
// Generate cache key from provider + voice + text + instructions
32
+
// Custom voice clones (jeffrey) get their own subfolder so all utterances
33
+
// from that voice are easy to list / browse on the CDN.
32
34
function getCacheKey(provider, voiceId, text, instructions) {
33
35
const parts = `${provider}:${voiceId}:${text}${instructions ? `:${instructions}` : ""}`;
34
36
const hash = crypto.createHash("sha256").update(parts).digest("hex");
35
-
return `${CACHE_PREFIX}${hash}.mp3`;
37
+
const subfolder = provider === "jeffrey" ? "jeffrey/" : "";
38
+
return `${CACHE_PREFIX}${subfolder}${hash}.mp3`;
36
39
}
37
40
38
41
// Check if cached audio exists, return CDN URL if so
···
49
52
}
50
53
}
51
54
52
-
// Save audio to cache
53
-
async function saveToCache(key, audioBuffer) {
55
+
// Save audio to cache. Optional `metadata` is persisted as S3 user metadata
56
+
// so we can browse utterances later (e.g. HeadObject → x-amz-meta-text).
57
+
async function saveToCache(key, audioBuffer, metadata = {}) {
54
58
try {
59
+
// S3 user metadata must be ASCII and each header is usually capped
60
+
// around 2 KB; trim text + encode non-ASCII defensively.
61
+
const cleanMeta = {};
62
+
for (const [k, v] of Object.entries(metadata)) {
63
+
if (v == null) continue;
64
+
const str = String(v).slice(0, 1800);
65
+
// Keep values ASCII-safe (S3 rejects high-unicode metadata headers).
66
+
cleanMeta[k] = Buffer.from(str, "utf8").toString("ascii").replace(/[\r\n]/g, " ");
67
+
}
68
+
55
69
await s3.send(new PutObjectCommand({
56
70
Bucket: BUCKET,
57
71
Key: key,
···
59
73
ContentType: "audio/mpeg",
60
74
ACL: "public-read",
61
75
CacheControl: "public, max-age=31536000", // 1 year (audio doesn't change)
76
+
Metadata: cleanMeta,
62
77
}));
63
78
console.log(`✅ Cached TTS: ${CDN_URL}/${key}`);
64
79
return `${CDN_URL}/${key}`;
···
147
162
};
148
163
}
149
164
165
+
// ── Jeffrey: Professional Voice Clone (PVC) ──────────────────────────
166
+
// Trained on multiple public lectures/talks by @jeffrey. Same voice
167
+
// used in the LACMA 2026 grant pitch video.
168
+
// Usage from the piece: `say:jeffrey hello world`
169
+
const JEFFREY_VOICE_ID = "dYNGZ848Oo6DtNBoeqgh";
170
+
171
+
async function generateJeffrey(text, scream) {
172
+
// Calmer, more natural delivery than the premade "scream" preset.
173
+
// Same knobs as the grant-video pipeline for homogeneity.
174
+
const voiceSettings = scream
175
+
? { stability: 0.2, similarity_boost: 0.85, style: 0.9, use_speaker_boost: true }
176
+
: { stability: 0.65, similarity_boost: 0.9, style: 0.15, use_speaker_boost: true };
177
+
178
+
const response = await fetch(
179
+
`https://api.elevenlabs.io/v1/text-to-speech/${JEFFREY_VOICE_ID}`,
180
+
{
181
+
method: "POST",
182
+
headers: {
183
+
"xi-api-key": process.env.ELEVENLABS_API_KEY,
184
+
"Content-Type": "application/json",
185
+
},
186
+
body: JSON.stringify({
187
+
text,
188
+
model_id: "eleven_multilingual_v2",
189
+
voice_settings: voiceSettings,
190
+
}),
191
+
},
192
+
);
193
+
194
+
if (!response.ok) {
195
+
const err = await response.text();
196
+
throw new Error(`ElevenLabs (Jeffrey) API error ${response.status}: ${err}`);
197
+
}
198
+
199
+
return {
200
+
buffer: Buffer.from(await response.arrayBuffer()),
201
+
voiceId: "jeffrey-pvc",
202
+
};
203
+
}
204
+
150
205
// Generate audio with Google Cloud TTS
151
206
async function generateGoogle(text, gender, set, isSSML) {
152
207
// Fetch GCP key from URL
···
271
326
result = await generateGoogle(text, gender, set, isSSML);
272
327
} else if (provider === "eleven") {
273
328
result = await generateElevenLabs(text, gender, set, scream);
329
+
} else if (provider === "jeffrey") {
330
+
result = await generateJeffrey(text, scream);
274
331
} else {
275
332
result = await generateOpenAI(text, gender, set, instructions);
276
333
}
···
287
344
288
345
console.log(`🗣️ Generated with ${provider}: ${voiceId}`);
289
346
290
-
// Cache for next time
291
-
const cdnUrl = await saveToCache(cacheKey, audioBuffer);
347
+
// Cache for next time. Attach the original text + voice as S3 metadata
348
+
// so individual objects are self-describing when you browse them.
349
+
const cdnUrl = await saveToCache(cacheKey, audioBuffer, {
350
+
text,
351
+
provider,
352
+
voice: voiceId,
353
+
scream: scream ? "1" : "0",
354
+
ts: new Date().toISOString(),
355
+
});
292
356
293
357
if (cdnUrl) {
294
358
return {
+22
-6
system/public/aesthetic.computer/disks/say.mjs
+22
-6
system/public/aesthetic.computer/disks/say.mjs
···
2
2
// A simple test piece for the TTS API.
3
3
// Type a word or phrase after `say` to hear it spoken.
4
4
5
-
/* #region 📚 README
5
+
/* #region 📚 README
6
6
Usage: say hello
7
7
say how are you today
8
8
say:male hi there
9
9
say:female good morning
10
10
say:google hello (use Google TTS)
11
11
say:google:female hi there
12
+
say:jeffrey hello (Jeffrey PVC — @jeffrey's voice clone)
13
+
say:jeffrey:scream AHHH (screaming variant)
14
+
15
+
All utterances are cached to the art.aesthetic.computer CDN under
16
+
`tts-cache/` (jeffrey gets its own subfolder `tts-cache/jeffrey/`).
17
+
Each cached MP3 has the original text + voice stamped as S3 metadata,
18
+
so you can rehydrate the catalog later by listing the bucket.
12
19
#endregion */
13
20
14
21
let text = "";
···
34
41
if (part === "google") provider = "google";
35
42
else if (part === "openai") provider = "openai";
36
43
else if (part === "eleven") provider = "eleven";
44
+
else if (part === "jeffrey") provider = "jeffrey";
37
45
else if (part === "male") gender = "male";
38
46
else if (part === "female") gender = "female";
39
47
else if (part === "scream") {
40
48
scream = true;
41
49
if (provider === "openai") {
42
50
instructions = "Deliver this as a blood-curdling scream. Shriek at the absolute top of your lungs with your voice cracking. Pure primal rage. Do NOT speak normally — only scream, raw and unhinged.";
43
-
} else if (provider !== "eleven") {
44
-
provider = "eleven"; // Default scream to ElevenLabs
51
+
} else if (provider !== "eleven" && provider !== "jeffrey") {
52
+
provider = "eleven"; // Default scream to ElevenLabs premade voices
45
53
}
46
54
}
47
55
}
···
56
64
// Note: Top-left corner is reserved for prompt HUD label
57
65
58
66
// Provider indicator (below HUD area)
59
-
const providerColor = scream ? "red" : provider === "eleven" ? "orange" : provider === "google" ? "cyan" : "lime";
67
+
const providerColor = scream
68
+
? "red"
69
+
: provider === "jeffrey"
70
+
? "magenta"
71
+
: provider === "eleven"
72
+
? "orange"
73
+
: provider === "google"
74
+
? "cyan"
75
+
: "lime";
60
76
const providerLabel = scream ? `[${provider} SCREAM]` : `[${provider}]`;
61
77
ink(providerColor).write(providerLabel, { x: 6, y: 18 });
62
-
78
+
63
79
// Instructions
64
80
ink("gray").write("say <words>", { x: 6, y: 32 });
65
-
ink("gray").write("say:google or say:male", { x: 6, y: 44 });
81
+
ink("gray").write("say:jeffrey · say:google · say:male", { x: 6, y: 44 });
66
82
67
83
// Current text
68
84
if (text) {