Aime/app/recorder.js

import { spawn, execFile } from 'node:child_process';
import { promisify } from 'node:util';
import fs from 'node:fs';
import path from 'node:path';
import readline from 'node:readline';
import * as diar from './diar.js';
import * as llm from './llm.js';
const pexec = promisify(execFile);

const WHISPER = '/opt/whisper.cpp/build/bin/whisper-cli';
// medium since 2026-07-08: user reported small mis-hears words on the USB-headset
// call audio (cleaner than the old BT-mSBC, so the extra accuracy now pays off).
// ~2x slower than small + more heat on long recordings (thermal throttle >85 C) —
// acceptable trade for accuracy. Override with WHISPER_MODEL to switch back.
const MODEL = process.env.WHISPER_MODEL || '/opt/whisper.cpp/models/ggml-medium.bin';
// NPU turbo transcriber: large-v3-turbo ENCODER on the RK3588 NPU (rknn fp16),
// decoder on CPU. Benchmarked ~2.4x faster than whisper.cpp medium on the R1
// (RTF 0.64 vs 1.51) and the encoder runs off-CPU -> far less A76 heat. The
// Python CLI emits {text, words:[{word,t_start_ms,t_end_ms}]}. Set TRANSCRIBER=cpu
// to fall back to whisper.cpp medium (the block below).
const TRANSCRIBER = process.env.TRANSCRIBER || 'npu';
const NPU_PY = '/opt/aime-venv/bin/python';
const NPU_CLI = '/opt/callrec/inference/transcribe_cli.py';
// SHORT clips (<= SHORT_SEC): a warm CPU daemon + resident rkllm instead of
// the NPU path, whose per-call model load (~28-40s) + rkllm swap (~40s) dwarf
// the actual work on a tiny clip. See the daemon client below.
const TD_DAEMON = '/opt/callrec/inference/transcribe_daemon_cpu.py';
const SHORT_SEC = +(process.env.SHORT_TRANSCRIBE_SEC || 30);
const REC_DIR = '/opt/callrec/recordings';
fs.mkdirSync(REC_DIR, { recursive: true });

// PipeWire/Pulse runs in the callrec user session; the systemd unit already
// sets XDG_RUNTIME_DIR + DBUS_SESSION_BUS_ADDRESS, so just inherit the env.
function penv() { return process.env; }

// --- Thermal safety for whisper on the fanless board -----------------------
// A 79-min recording transcribed with ggml-medium pinned to all 4 A76 cores at
// max clock drove the big cores to CRITICAL temp -> the kernel did a HARDWARE
// PROTECTION shutdown (auto-reboot), losing the in-flight recording. Two guards:
//   1. cap the A76 max clock while transcribing (biggest heat lever), and
//   2. a hard backstop that SIGSTOPs whisper if the big cores still cross 90 C
//      and SIGCONTs it once they fall back under 78 C.
const A76_MAX = 2256000;   // restore-to clock (from scaling_available_frequencies)
const A76_CAP = 1608000;   // 1.6 GHz cap during whisper
const A76_CAP_NPU = 1800000; // relaxed 1.8 GHz for the NPU decoder (encoder is off-CPU;
                            // measured only ~2% slower at 1.6 but ~20 C cooler than full 2.25)
const TEMP_ZONES = ['/sys/class/thermal/thermal_zone1/temp',   // bigcore0
                    '/sys/class/thermal/thermal_zone2/temp'];  // bigcore1
const TEMP_PAUSE = 90000;  // >90 C -> pause whisper
const TEMP_RESUME = 78000; // <78 C -> resume

function readMaxBigTemp() {
  let m = 0;
  for (const z of TEMP_ZONES) {
    try { const v = parseInt(fs.readFileSync(z, 'utf8'), 10); if (v > m) m = v; } catch {}
  }
  return m;
}
// callrec has a NOPASSWD sudo rule for exactly /opt/callrec/aime-cpucap.sh.
function setA76Max(freq) {
  return new Promise(res => execFile('sudo', ['-n', '/opt/callrec/aime-cpucap.sh', String(freq)], () => res()));
}

// Spawn whisper on the A76 cores with the SIGSTOP/SIGCONT temp backstop. The A76
// CLOCK CAP is NOT managed here — the caller (transcribeWav) holds it across the
// WHOLE heavy pipeline (whisper + diarization), because diarization also pins the
// A76 cores and crashed the board when it ran at full clock after whisper's cap
// was released. Rejects on non-zero exit.
async function runWhisperGuarded(args) {
  let guard = null, paused = false;
  try {
    await new Promise((resolve, reject) => {
      const p = spawn('taskset', ['-c', '4-7', WHISPER, ...args], { stdio: ['ignore', 'ignore', 'pipe'] });
      let err = '';
      p.stderr.on('data', d => { err += d; if (err.length > (1 << 20)) err = err.slice(-(1 << 19)); });
      guard = setInterval(() => {
        const t = readMaxBigTemp();
        if (!paused && t >= TEMP_PAUSE) {
          try { process.kill(p.pid, 'SIGSTOP'); paused = true; console.warn(`whisper PAUSED (A76 ${t / 1000} C)`); } catch {}
        } else if (paused && t > 0 && t <= TEMP_RESUME) {
          try { process.kill(p.pid, 'SIGCONT'); paused = false; console.warn(`whisper resumed (A76 ${t / 1000} C)`); } catch {}
        }
      }, 3000);
      p.on('error', reject);
      p.on('close', code => code === 0 ? resolve() : reject(new Error('whisper exit ' + code + ': ' + err.slice(-400))));
    });
  } finally {
    if (guard) clearInterval(guard);
  }
}

// Run the NPU turbo transcriber (Python) under the same A76 temp backstop as
// whisper. taskset execs python -> p.pid IS python, so SIGSTOP/SIGCONT reach it.
// Returns the parsed CLI JSON. The A76 CLOCK CAP is held by the caller.
async function runNpuTranscribe(wavPath) {
  let guard = null, paused = false, out = '', err = '';
  try {
    await new Promise((resolve, reject) => {
      const p = spawn('taskset', ['-c', '4-7', NPU_PY, NPU_CLI, wavPath],
        { stdio: ['ignore', 'pipe', 'pipe'], cwd: '/opt/callrec' });
      p.stdout.on('data', d => { out += d; });
      p.stderr.on('data', d => { err += d; if (err.length > (1 << 20)) err = err.slice(-(1 << 19)); });
      guard = setInterval(() => {
        const t = readMaxBigTemp();
        if (!paused && t >= TEMP_PAUSE) {
          try { process.kill(p.pid, 'SIGSTOP'); paused = true; console.warn(`transcribe PAUSED (A76 ${t / 1000} C)`); } catch {}
        } else if (paused && t > 0 && t <= TEMP_RESUME) {
          try { process.kill(p.pid, 'SIGCONT'); paused = false; console.warn(`transcribe resumed (A76 ${t / 1000} C)`); } catch {}
        }
      }, 3000);
      p.on('error', reject);
      p.on('close', code => code === 0 ? resolve() : reject(new Error('npu-transcribe exit ' + code + ': ' + err.slice(-400))));
    });
  } finally {
    if (guard) clearInterval(guard);
  }
  const line = out.trim();
  if (!line) throw new Error('npu-transcribe: leere Ausgabe. stderr: ' + err.slice(-400));
  return JSON.parse(line);
}

// --- Warm CPU transcriber daemon (SHORT clips) -----------------------------
// large-v3-turbo CT2 on the CPU, loaded ONCE and kept warm. CPU-only, so it
// coexists with a resident rkllm_server on the NPU (no memory-wall conflict):
// a short clip needs neither a per-call model load nor an rkllm swap. Lazy-
// spawned on the first short clip, then persistent. base64 line protocol.
let tdProc = null, tdReady = null, tdRl = null, tdPending = null, tdChain = Promise.resolve();
function startTd() {
  if (tdProc) return tdReady;
  tdProc = spawn('taskset', ['-c', '4-7', NPU_PY, TD_DAEMON],
    { stdio: ['pipe', 'pipe', 'pipe'], cwd: '/opt/callrec', env: { ...process.env, WHISPER_BATCH: process.env.WHISPER_BATCH || '1' } });
  tdProc.stderr.on('data', () => {}); // onnxruntime GPU-probe warnings, ignore
  tdReady = new Promise((res, rej) => {
    const to = setTimeout(() => rej(new Error('transcribe-daemon READY timeout')), 120000);
    tdRl = readline.createInterface({ input: tdProc.stdout });
    tdRl.on('line', line => {
      if (line.startsWith('READY')) { clearTimeout(to); res(); return; }
      if (!tdPending) return;
      const p = tdPending; tdPending = null;
      if (line.startsWith('OK ')) { try { p.resolve(JSON.parse(Buffer.from(line.slice(3), 'base64').toString('utf8'))); } catch (e) { p.reject(e); } }
      else p.reject(new Error(Buffer.from(line.slice(4), 'base64').toString('utf8') || 'daemon error'));
    });
  });
  const cleanup = () => { tdProc = null; tdReady = null; tdRl = null; if (tdPending) { tdPending.reject(new Error('transcribe-daemon beendet')); tdPending = null; } };
  tdProc.on('exit', cleanup); tdProc.on('error', cleanup);
  return tdReady;
}
// One request at a time (single warm process), chained like spk.embed.
function runDaemonTranscribe(wavPath) {
  tdChain = tdChain.then(async () => {
    await startTd();
    return await new Promise((res, rej) => { tdPending = { resolve: res, reject: rej }; tdProc.stdin.write(wavPath + '\n'); });
  });
  return tdChain;
}
// Wav duration from file size (16 kHz mono s16 = 32000 B/s, 44 B header).
function wavDurationSec(wavPath) {
  try { return Math.max(0, (fs.statSync(wavPath).size - 44) / 32000); } catch { return 0; }
}

let current = null; // { proc, wavPath, startTs, source, title }

export function status() {
  return current ? { recording: true, startedAt: current.startTs, source: current.source } : { recording: false };
}
export function isRecording() { return !!current; }

// List real capture devices (one entry per recording device). A "source" in
// PulseAudio/PipeWire includes monitor sources (the loopback of every OUTPUT)
// and both A2DP/HFP variants of a BT card — confusing for the user. We only
// want actual microphones/inputs, labelled by their friendly device name.
export async function listSources() {
  try {
    const { stdout } = await pexec('pactl', ['list', 'sources'], { env: penv() });
    const out = [];
    const seen = new Set();
    for (const block of stdout.split(/\nSource #/).map(b => b.trim()).filter(Boolean)) {
      const name = (block.match(/^\s*Name:\s*(.+)$/m) || [])[1];
      if (!name) continue;
      // Skip monitors (loopback of an output) — not a recording device.
      // Every block has a "Monitor of Sink:" line; a real input reads "n/a".
      if (/\.monitor$/i.test(name) || /Monitor of Sink:\s*(?!n\/a)\S/i.test(block)) continue;
      const isBluetooth = /bluez/i.test(name);
      // For BT the physical device is one card; both bluez_input.* names map to
      // the same headset, so collapse by the MAC to a single entry.
      const key = isBluetooth ? (name.match(/([0-9A-F]{2}[:_]){5}[0-9A-F]{2}/i) || [name])[0] : name;
      if (seen.has(key)) continue;
      seen.add(key);
      const label = (block.match(/^\s*Description:\s*(.+)$/m) || [])[1] || name;
      out.push({ name, label, isBluetooth });
    }
    return out;
  } catch (e) { return []; }
}

// Find the bluez card name (e.g. bluez_card.98_34_8C_FD_24_17).
async function bluezCard() {
  const { stdout } = await pexec('pactl', ['list', 'cards', 'short'], { env: penv() });
  return stdout.split('\n').map(l => l.split('\t')[1]).find(n => /bluez/i.test(n || ''));
}

// List available profiles of the bluez card (for diagnostics / picking HFP).
async function cardProfiles(card) {
  const { stdout } = await pexec('pactl', ['list', 'cards'], { env: penv() });
  const block = stdout.split('Card #').find(b => b.includes(card));
  if (!block) return [];
  const m = block.match(/Profiles:[\s\S]*?(?=\n\tActive Profile:|\n\tPorts:|$)/);
  return m ? [...m[0].matchAll(/^\t\t([\w-]+):/gm)].map(x => x[1]) : [];
}

// Switch the connected BT headset to an HFP profile (bidirectional, mic-capable).
// A2DP has no microphone, so recording a call REQUIRES HFP (headset-head-unit).
// codec: 'msbc' (16 kHz wideband), 'cvsd' (8 kHz narrowband), or 'auto' (mSBC first).
// On this card: headset-head-unit = mSBC, headset-head-unit-cvsd = CVSD.
export async function enableHeadsetMic(codec = 'auto') {
  try {
    const card = await bluezCard();
    if (!card) return { ok: false, msg: 'no bluez card (headset connected?)' };
    const avail = await cardProfiles(card);
    const lists = {
      // mSBC = the plain "headset-head-unit" here (some stacks name it -msbc).
      msbc: ['headset-head-unit-msbc', 'headset-head-unit', 'handsfree-head-unit-msbc', 'handsfree-head-unit'],
      cvsd: ['headset-head-unit-cvsd', 'handsfree-head-unit-cvsd'],
      auto: ['headset-head-unit-msbc', 'headset-head-unit', 'handsfree-head-unit-msbc',
             'handsfree-head-unit', 'headset-head-unit-cvsd', 'handsfree-head-unit-cvsd'],
    };
    const prefer = lists[codec] || lists.auto;
    const pick = prefer.find(p => avail.includes(p));
    if (!pick) return { ok: false, msg: `codec ${codec} nicht verfügbar (${avail.join(', ')})`, available: avail };
    await pexec('pactl', ['set-card-profile', card, pick], { env: penv() });
    return { ok: true, card, profile: pick, codec, available: avail };
  } catch (e) { return { ok: false, msg: String(e && e.message || e) }; }
}

export async function startRecording({ source, title, codec, numSpeakers } = {}) {
  if (current) throw new Error('already recording');
  // Bluetooth headset: force HFP so the microphone actually delivers audio
  // (A2DP is playback-only). Do this right before capture; the profile holds
  // while the source is open. codec = 'msbc' | 'cvsd' | 'auto' (A/B testing).
  let hfp = null;
  if (source && /bluez/i.test(source)) {
    hfp = await enableHeadsetMic(codec || 'auto');
    if (!hfp.ok) throw new Error('HFP-Profil konnte nicht aktiviert werden: ' + hfp.msg);
    // the source object is recreated under the new profile; wait, then re-resolve
    // the actual bluez input source name (it may differ from the A2DP-era name).
    await new Promise(r => setTimeout(r, 900));
    const srcs = await listSources();
    const bz = srcs.find(s => s.isBluetooth && !/monitor/i.test(s.name));
    if (bz) source = bz.name;
  }
  const ts = Date.now();
  const wavPath = path.join('/tmp', `rec_${ts}.wav`);
  // 16 kHz mono s16 = whisper's native format. NOTE: parecord takes the output
  // filename as a POSITIONAL arg (there is no --file= option).
  const args = ['--file-format=wav', '--rate=16000', '--channels=1', '--format=s16le'];
  if (source) args.push(`--device=${source}`);
  args.push(wavPath);
  const proc = spawn('parecord', args, { env: penv() });
  let err = '';
  proc.stderr.on('data', d => { err += d; });
  proc.on('error', e => { err += String(e); });
  // Label the recording with the actual codec used so A/B tests are obvious
  // in the list (mSBC 16 kHz vs CVSD 8 kHz).
  const codecTag = hfp && hfp.profile ? (/cvsd/i.test(hfp.profile) ? ' [CVSD 8k]' : ' [mSBC 16k]') : '';
  current = {
    proc, wavPath, startTs: ts, source: source || 'default', numSpeakers,
    title: (title || `Aufnahme ${new Date(ts).toLocaleString('de-DE')}`) + codecTag, err: () => err
  };
  return { startedAt: ts };
}

// Stop capture and return the raw wav path only (no whisper/mp3/diarization).
// Used for voiceprint enrollment.
export async function stopRaw() {
  if (!current) throw new Error('not recording');
  const c = current; current = null;
  c.proc.kill('SIGINT');
  await new Promise(resolve => {
    if (c.proc.exitCode !== null || c.proc.signalCode !== null) return resolve();
    let done = false; const fin = () => { if (!done) { done = true; resolve(); } };
    c.proc.on('close', fin); setTimeout(fin, 4000);
  });
  if (!fs.existsSync(c.wavPath) || fs.statSync(c.wavPath).size < 1000)
    throw new Error('Aufnahme leer/fehlgeschlagen. parecord: ' + c.err());
  return c.wavPath;
}

// Stop capture and PERSIST the wav + mp3 to REC_DIR immediately — BEFORE any
// transcription. This is the crash-safety fix: whisper on a long clip can crash
// the board (thermal), so the audio must already be on rootfs (not /tmp) and the
// DB row created before whisper runs. Call transcribeWav() afterwards (async).
export async function stopAndSave() {
  if (!current) throw new Error('not recording');
  const c = current; current = null;
  c.proc.kill('SIGINT');
  await new Promise(resolve => {
    if (c.proc.exitCode !== null || c.proc.signalCode !== null) return resolve();
    let done = false;
    const fin = () => { if (!done) { done = true; resolve(); } };
    c.proc.on('close', fin);
    setTimeout(fin, 4000);
  });
  const dur = Date.now() - c.startTs;
  if (!fs.existsSync(c.wavPath) || fs.statSync(c.wavPath).size < 1000)
    throw new Error('Aufnahme leer/fehlgeschlagen. parecord: ' + c.err());

  const base = `rec_${c.startTs}`;
  const mp3Path = path.join(REC_DIR, `${base}.mp3`);
  // Gentle cleanup for playback: cut low rumble + normalize loudness. Applied to
  // the MP3 only; whisper still transcribes the untouched WAV.
  await pexec('ffmpeg', ['-y', '-i', c.wavPath,
    '-af', 'highpass=f=80,loudnorm=I=-16:TP=-1.5:LRA=11',
    '-codec:a', 'libmp3lame', '-qscale:a', '3', mp3Path]);

  // Move the wav into REC_DIR (rootfs — survives a reboot); it is NOT left in
  // /tmp, so a crash during transcription can no longer lose it.
  const wavFile = `${base}.wav`;
  try { fs.renameSync(c.wavPath, path.join(REC_DIR, wavFile)); }
  catch { fs.copyFileSync(c.wavPath, path.join(REC_DIR, wavFile)); try { fs.unlinkSync(c.wavPath); } catch {} }

  return { mp3_file: `${base}.mp3`, wav_file: wavFile, duration_ms: dur,
           language: 'de', title: c.title, device: c.source, numSpeakers: c.numSpeakers };
}

// Transcribe a saved wav (in REC_DIR) with whisper (thermal-guarded) + diarize.
// Returns { transcript, words }. Idempotent — safe to re-run for recovery of a
// recording whose transcription was interrupted (crash) or failed.
export async function transcribeWav(wavFile, numSpeakers) {
  const wavPath = path.isAbsolute(wavFile) ? wavFile : path.join(REC_DIR, wavFile);
  if (!fs.existsSync(wavPath)) throw new Error('wav-Datei fehlt: ' + wavFile);
  const outBase = `/tmp/${path.basename(wavPath, '.wav')}`;
  const words = [];
  let transcript = '';
  // Diarization (sherpa-onnx over the whole recording) pins all 4 A76 cores and
  // crashed the board at full clock on a 46-min call, so it ALWAYS runs under the
  // A76 clock cap. Transcription is capped ONLY on the whisper.cpp (medium) path
  // (CPU-hot); the NPU-turbo path runs the encoder on the NPU (CPU only decodes,
  // ~59-69 C) so it needs NO cap. rkllm_server (~3.1 GB) is freed first on the NPU
  // path (turbo rknn + CT2 decoder OOM-killed next to it); it respawns for analysis.
  const wavSec = wavDurationSec(wavPath);
  try {
    if (TRANSCRIBER === 'npu' && wavSec > 0 && wavSec <= SHORT_SEC) {
      // SHORT clip: warm CPU daemon + keep rkllm RESIDENT (no NPU swap), at
      // full clock (a few seconds of compute won't overheat the fanless board).
      // Cuts a 10s clip from ~120s (per-call model load + rkllm swap) to ~30s.
      const r = await runDaemonTranscribe(wavPath);
      for (const w of r.words || []) {
        const t = (w.word || '').trim();
        if (!t) continue;
        words.push({ word: t, t_start_ms: w.t_start_ms, t_end_ms: w.t_end_ms });
        transcript += (transcript ? ' ' : '') + t;
      }
      console.log(`cpu-daemon-transcribe: ${words.length} words, RTF ${r.rtf} (${r.transcribe_s}s / ${r.duration}s audio, warm, rkllm kept resident)`);
    } else if (TRANSCRIBER === 'npu') {
      await llm.stopServer();
      await setA76Max(A76_CAP_NPU);
      // large-v3-turbo, encoder on NPU. CLI already emits word-level timestamps.
      const r = await runNpuTranscribe(wavPath);
      for (const w of r.words || []) {
        const t = (w.word || '').trim();
        if (!t) continue;
        words.push({ word: t, t_start_ms: w.t_start_ms, t_end_ms: w.t_end_ms });
        transcript += (transcript ? ' ' : '') + t;
      }
      console.log(`npu-transcribe: ${words.length} words, RTF ${r.rtf} (load ${r.load_s}s + transcribe ${r.transcribe_s}s / ${r.duration}s audio)`);
    } else {
      await setA76Max(A76_CAP);
      // Fallback: whisper.cpp medium. Greedy (-bs 1 -bo 1) ~= beam-5 on German,
      // ~30% faster + less heat. whisper also has a SIGSTOP temp guard.
      await runWhisperGuarded(['-m', MODEL, '-f', wavPath, '-l', 'de', '-t', '4',
        '-bs', '1', '-bo', '1', '-ml', '1', '-sow', '-oj', '-of', outBase]);
      const json = JSON.parse(fs.readFileSync(`${outBase}.json`, 'utf8'));
      for (const seg of json.transcription || []) {
        const t = (seg.text || '').trim();
        if (!t) continue;
        words.push({ word: t, t_start_ms: seg.offsets.from, t_end_ms: seg.offsets.to });
        transcript += (transcript ? ' ' : '') + t;
      }
    }
    // Acoustic diarization under the A76 clock cap (both paths). Non-fatal.
    await setA76Max(A76_CAP);
    try {
      const segs = await diar.diarize(wavPath, numSpeakers);
      diar.assignSpeakers(words, segs);
    } catch (e) { console.error('diarize failed:', e.message); }
    try { fs.unlinkSync(`${outBase}.json`); } catch {}
  } finally {
    await setA76Max(A76_MAX);
  }
  return { transcript, words };
}