// kale-tts self-hosted HTTP API (OpenAI-shaped). POST /v1/audio/speech {model, input, voice, response_format: "wav"|"pcm", stream} // model: kale-tts-1-nano (default) | kale-tts-1-tiny; voice by name: kenzie simone amy frankie darlene owen toby jack earl magnus (GET /v1/voices) // stream: true (or response_format "pcm") -> chunked raw 16-bit 24 kHz mono PCM, one chunk per sentence as it is ready // otherwise one WAV. With TTS_API_TOKEN set (>= 20 chars) every route except /health needs "Authorization: Bearer "; // without it the server only listens on 127.0.0.1 (local use). // run (in the release folder or a downloaded model repo): npm i onnxruntime-node && node server/server.mjs (env: PORT=8000, HOST, KALE_TTS_WORKERS=, KALE_TTS_THREADS=1, // KALE_TTS_ROOT=/) import http from "node:http"; import { timingSafeEqual, createHash } from "node:crypto"; import cluster from "node:cluster"; import os from "node:os"; // onnxruntime-node runs a session on the JS thread, so one process renders one request at a time (a session pool doesn't help). // Parallel requests: KALE_TTS_WORKERS processes on the same port (node:cluster), default = CPU cores (max 8), 1 onnxruntime thread each. // Each process loads both models (~0.3 GB). Measured on a Mac: 1 process nano 152x, tiny 82x real time (18 Nov); 10 processes 465 s of audio/s (16 Nov) // with 10 parallel clients. const WORKERS = Number(process.env.KALE_TTS_WORKERS ?? Math.min(8, os.availableParallelism?.() ?? os.cpus().length)); if (cluster.isPrimary && WORKERS > 1) { for (let i = 0; i < WORKERS; i++) cluster.fork(); cluster.on("exit", (w, code) => { if (code !== 0) { console.error(`kale-tts worker ${w.process.pid} exited (${code}), restarting`); cluster.fork(); } }); await new Promise(() => {}); } const T0 = performance.now(); const ROOT = process.env.KALE_TTS_ROOT ?? new URL("../", import.meta.url).pathname; // the release folder (js/, models/, frontend/) or one downloaded model repo const { KaleTTS, toWav } = await import(ROOT + "js/kale-tts.mjs"); const ort = await import("onnxruntime-node"); const threads = Number(process.env.KALE_TTS_THREADS ?? 1); // a downloaded model repo (voices.json at its root) serves that one model; the release folder serves both const { existsSync } = await import("node:fs"); const ONE = existsSync(ROOT + "voices.json") ? (process.env.KALE_TTS_MODEL ?? (existsSync(ROOT + "ettin-tokenizer.json") ? "kale-tts-1-tiny" : "kale-tts-1-nano")) : null; const MODELS = ONE ? [ONE] : ["kale-tts-1-nano", "kale-tts-1-tiny"]; const tts = {}; for (const m of MODELS) tts[m] = await KaleTTS.load(ONE ? ROOT : m, { ort, threads }); for (const m of MODELS) { await tts[m].preload(); for (const v of tts[m].listVoices()) await tts[m].speak("Hi.", { voice: v.id }); } // load voices + warm const TOKEN = process.env.TTS_API_TOKEN ?? ""; if (TOKEN && TOKEN.length < 20) throw new Error("TTS_API_TOKEN must be at least 20 characters"); const h = (s) => createHash("sha256").update(s).digest(); const authed = (req) => { if (!TOKEN) return true; const a = req.headers.authorization ?? ""; return a.startsWith("Bearer ") && timingSafeEqual(h(a.slice(7)), h(TOKEN)); }; const json = (res, code, obj) => { res.writeHead(code, { "content-type": "application/json" }); res.end(JSON.stringify(obj)); }; http.createServer(async (req, res) => { try { if (req.method === "GET" && req.url === "/health") return res.end("ok"); if (!authed(req)) return json(res, 401, { error: "missing or wrong bearer token" }); if (req.method === "GET" && req.url === "/v1/models") return json(res, 200, { data: MODELS.map((id) => ({ id, object: "model" })) }); if (req.method === "GET" && req.url === "/v1/voices") return json(res, 200, { data: tts[MODELS[0]].listVoices() }); if (req.method !== "POST" || req.url !== "/v1/audio/speech") return json(res, 404, { error: "not found" }); let body = ""; for await (const c of req) { body += c; if (body.length > 20000) return json(res, 413, { error: "input too long (max ~20k chars)" }); } const b = JSON.parse(body || "{}"); const input = String(b.input ?? "").trim(); if (!input) return json(res, 400, { error: "input is empty" }); const model = MODELS.includes(b.model) ? b.model : MODELS[0]; const t = tts[model]; let voice; try { voice = t.resolveVoice(b.voice ?? t.listVoices()[0].id); } catch { return json(res, 400, { error: `unknown voice ${b.voice}`, voices: t.listVoices().map((v) => ({ name: (v.name ?? v.id).toLowerCase(), id: v.id })) }); } // public name (any case) or id const t0 = performance.now(); if (b.stream || b.response_format === "pcm") { res.writeHead(200, { "content-type": "audio/pcm", "x-sample-rate": "24000", "x-channels": "1", "x-sample-format": "s16le", "cache-control": "no-store" }); for await (const c of t.stream(input, { voice })) res.write(toWav(c.audio).subarray(44)); return res.end(); } const { audio } = await t.speak(input, { voice }); const w = toWav(audio); res.writeHead(200, { "content-type": "audio/wav", "content-length": w.length, "x-compute-ms": Math.round(performance.now() - t0) }); res.end(w); } catch (e) { if (!res.headersSent) json(res, 400, { error: String(e?.message ?? e) }); else res.end(); } }).listen(Number(process.env.PORT ?? 8000), process.env.HOST ?? (TOKEN ? "0.0.0.0" : "127.0.0.1"), () => console.log(`kale-tts api ready after ${((performance.now() - T0) / 1000).toFixed(2)} s`));