@inetafrica/open-claudia 3.5.47 → 3.5.59

package/channels/kazee/call/miss-capture.js

@@ -0,0 +1,82 @@
+"use strict";
+
+// Segments the recogniser returned no words for. On 12 Oct a caller's question
+// came back as nine empty finals in a row (1.2-3.7 s of audio each) and the
+// call went silent; nothing showed whether that audio was quiet, garbled or
+// clean speech. Every final now logs how loud its audio was, and with
+// KAZEE_CALL_KEEP_MISSES=on the empty ones are kept as WAV for a day.
+
+const fs = require("fs");
+const path = require("path");
+const { resolveConfigDir } = require("../../../core/config-dir");
+const { STT_RATE, rmsDbfs } = require("./audio");
+
+const FRAME_BYTES = (STT_RATE / 50) * 2; // 20 ms of 16-bit mono
+const MAX_SEGMENT_BYTES = STT_RATE * 2 * 35; // the turn cap is 30 s
+const MIN_KEEP_BYTES = STT_RATE * 2; // under 1 s is a cough, not a question
+const KEEP_FILES = 40;
+const KEEP_MS = 24 * 60 * 60_000;
+
+// The bot's settings: ~/.open-claudia/.env is loaded into core/config's
+// config, not into process.env (3.5.50 read process.env, so the switch in
+// .env never took effect).
+function settings() {
+  try { return require("../../../core/config").config || process.env; } catch (_) { return process.env; }
+}
+
+function missDir(env = settings()) {
+  return /^(on|1|true|yes)$/i.test(String(env.KAZEE_CALL_KEEP_MISSES || "").trim())
+    ? path.join(resolveConfigDir(env), "call-misses")
+    : null;
+}
+
+// rms: the whole segment; loud: its 90th-percentile 20 ms frame, i.e. how
+// loud the voice in it was. Both dBFS, rounded.
+function segmentLevel(pcm16) {
+  if (!pcm16?.length) return { rms: -Infinity, loud: -Infinity };
+  const frames = [];
+  for (let at = 0; at + FRAME_BYTES <= pcm16.length; at += FRAME_BYTES) {
+    frames.push(rmsDbfs(pcm16.subarray(at, at + FRAME_BYTES)));
+  }
+  frames.sort((a, b) => a - b);
+  const loud = frames.length ? frames[Math.min(frames.length - 1, Math.floor(frames.length * 0.9))] : -Infinity;
+  return { rms: Math.round(rmsDbfs(pcm16)), loud: Math.round(loud) };
+}
+
+function wav(pcm16, rate = STT_RATE) {
+  const header = Buffer.alloc(44);
+  header.write("RIFF", 0);
+  header.writeUInt32LE(36 + pcm16.length, 4);
+  header.write("WAVEfmt ", 8);
+  header.writeUInt32LE(16, 16);
+  header.writeUInt16LE(1, 20);
+  header.writeUInt16LE(1, 22);
+  header.writeUInt32LE(rate, 24);
+  header.writeUInt32LE(rate * 2, 28);
+  header.writeUInt16LE(2, 32);
+  header.writeUInt16LE(16, 34);
+  header.write("data", 36);
+  header.writeUInt32LE(pcm16.length, 40);
+  return Buffer.concat([header, pcm16]);
+}
+
+function prune(dir, now = Date.now()) {
+  const files = fs.readdirSync(dir).filter(name => name.endsWith(".wav"))
+    .map(name => ({ file: path.join(dir, name), mtime: fs.statSync(path.join(dir, name)).mtimeMs }))
+    .sort((a, b) => b.mtime - a.mtime);
+  files.forEach(({ file, mtime }, index) => {
+    if (index >= KEEP_FILES || now - mtime > KEEP_MS) fs.rmSync(file, { force: true });
+  });
+}
+
+// Returns the kept file, or null when the segment is too short to matter.
+function keepMiss(dir, name, pcm16) {
+  if (!dir || !pcm16 || pcm16.length < MIN_KEEP_BYTES) return null;
+  fs.mkdirSync(dir, { recursive: true, mode: 0o700 });
+  const file = path.join(dir, `${String(name).replace(/[^\w.-]+/g, "_")}.wav`);
+  fs.writeFileSync(file, wav(pcm16), { mode: 0o600 });
+  prune(dir);
+  return file;
+}
+
+module.exports = { MAX_SEGMENT_BYTES, MIN_KEEP_BYTES, KEEP_FILES, missDir, segmentLevel, keepMiss, wav, settings };