Live_RC/server/tts.js

// Local neural text-to-speech via Piper (https://github.com/rhasspy/piper). Fully offline:
// each spoken line is generated once, cached as a WAV under data/tts/, and replayed from
// disk after that. Works with the Python `piper-tts` package (Linux/macOS/Windows) and the
// standalone piper/piper.exe release - both accept -m, -f and --length_scale.
const crypto = require("crypto");
const fs = require("fs");
const os = require("os");
const path = require("path");
const { execFile } = require("child_process");

const MAX_TEXT_LENGTH = 300;
const GENERATE_TIMEOUT_MS = 20000;

// Order: explicit RC_PIPER_BIN, then the copy `npm run tts:install` puts in <project>/piper/
// (known-good standalone build), then PATH, then pipx's default bin dir - a background
// service often runs without the user's full PATH.
function resolvePiperBinary(configured, projectDir) {
  const names = process.platform === "win32" ? ["piper.exe", "piper.cmd", "piper"] : ["piper"];
  const candidates = [];
  if (configured) candidates.push(configured);
  names.forEach((name) => candidates.push(path.join(projectDir, "piper", name)));
  String(process.env.PATH || "").split(path.delimiter).filter(Boolean).forEach((dir) => names.forEach((name) => candidates.push(path.join(dir, name))));
  names.forEach((name) => candidates.push(path.join(os.homedir(), ".local", "bin", name)));
  return candidates.find((candidate) => {
    try { return fs.statSync(candidate).isFile(); } catch { return false; }
  }) || null;
}

// A voice is a model file plus its matching .onnx.json config, named <lang>_<REGION>-<name>-<quality>.
function listVoices(voiceDirs) {
  const voices = new Map();
  voiceDirs.forEach((dir) => {
    let entries = [];
    try { entries = fs.readdirSync(dir); } catch { return; }
    entries.filter((entry) => entry.endsWith(".onnx") && entries.includes(`${entry}.json`)).forEach((entry) => {
      const name = entry.slice(0, -".onnx".length);
      if (!voices.has(name)) voices.set(name, { name, file: path.join(dir, entry), language: name.split(/[_-]/)[0].toLowerCase() });
    });
  });
  return [...voices.values()].sort((a, b) => a.name.localeCompare(b.name));
}

function pickVoice(voices, language) {
  const prefix = String(language || "").toLowerCase().split("-")[0];
  const matching = voices.filter((voice) => voice.language === prefix);
  return matching.find((voice) => voice.name.endsWith("-medium")) || matching[0] || null;
}

function createTts({ dataDir, projectDir, piperBin = process.env.RC_PIPER_BIN || "", voiceDirs = [], run = execFile }) {
  const cacheDir = path.join(dataDir, "tts");
  const inFlight = new Map();
  let queue = Promise.resolve();

  function status() {
    const binary = resolvePiperBinary(piperBin, projectDir);
    const voices = listVoices(voiceDirs);
    let cachedLines = 0;
    try { cachedLines = fs.readdirSync(cacheDir).filter((entry) => entry.endsWith(".wav")).length; } catch {}
    return { available: Boolean(binary && voices.length), binary, voices: voices.map(({ name, language }) => ({ name, language })), voiceDirs, cachedLines };
  }

  // The voice name and speed are part of the key, so changing either never replays a stale line.
  function cachePathFor(voice, text, lengthScale) {
    const key = crypto.createHash("sha256").update(`${voice.name}\n${lengthScale}\n${text}`).digest("hex");
    return path.join(cacheDir, `${key}.wav`);
  }

  // One Piper process at a time: loading a model is CPU-heavy, and a burst of parallel
  // generations must never starve the timing server itself.
  function generate(binary, voice, text, lengthScale, outPath) {
    const job = queue.then(() => new Promise((resolve, reject) => {
      fs.mkdirSync(cacheDir, { recursive: true });
      const tmpPath = `${outPath}.${process.pid}.tmp`;
      const child = run(binary, ["-m", voice.file, "-f", tmpPath, "--length_scale", String(lengthScale)], { timeout: GENERATE_TIMEOUT_MS, windowsHide: true }, (error) => {
        if (error || !fs.existsSync(tmpPath)) {
          fs.rmSync(tmpPath, { force: true });
          reject(error || new Error("Piper produced no audio"));
          return;
        }
        fs.renameSync(tmpPath, outPath);
        resolve(outPath);
      });
      child.stdin?.end(`${text}\n`);
    }));
    queue = job.catch(() => {});
    return job;
  }

  // Resolves to a cached WAV path, generating it first if needed. Throws with
  // code "unavailable" when Piper or a voice for the language is missing.
  async function speech({ text, language, rate }) {
    const cleanText = String(text || "").replace(/\s+/g, " ").trim().slice(0, MAX_TEXT_LENGTH);
    if (!cleanText) throw Object.assign(new Error("No text"), { code: "bad_request" });
    const binary = resolvePiperBinary(piperBin, projectDir);
    const voice = pickVoice(listVoices(voiceDirs), language);
    if (!binary || !voice) throw Object.assign(new Error("Piper or a voice for this language is not installed"), { code: "unavailable" });
    // Piper's length scale is duration, the inverse of the browser's speaking rate.
    const lengthScale = Math.round(Math.max(0.6, Math.min(1.6, 1 / (Number(rate) || 1))) * 100) / 100;
    const outPath = cachePathFor(voice, cleanText, lengthScale);
    if (fs.existsSync(outPath)) return outPath;
    if (!inFlight.has(outPath)) {
      inFlight.set(outPath, generate(binary, voice, cleanText, lengthScale, outPath).finally(() => inFlight.delete(outPath)));
    }
    return inFlight.get(outPath);
  }

  const progress = { running: false, total: 0, done: 0, failed: 0 };
  async function pregenerate(items) {
    if (progress.running) return progress;
    const list = (Array.isArray(items) ? items : []).filter((item) => String(item?.text || "").trim()).slice(0, 2000);
    Object.assign(progress, { running: true, total: list.length, done: 0, failed: 0 });
    for (const item of list) {
      try { await speech(item); } catch { progress.failed += 1; }
      progress.done += 1;
    }
    progress.running = false;
    return progress;
  }

  return { status: () => ({ ...status(), pregenerate: { ...progress } }), speech, pregenerate };
}

module.exports = { createTts, listVoices, pickVoice, resolvePiperBinary, MAX_TEXT_LENGTH };