@gotcos/glasses-server 6.37.1 → 6.37.2

package/CHANGELOG.md

@@ -1,3 +1,34 @@
+## 6.37.2
+
+`--setup-speaker-model` — one command to install the voiceprint model.
+
+Named per-speaker diarization needs a ~26 MB model that is deliberately not in
+the npm tarball. Until now the only way to get it was the README telling you to
+"put the .onnx there", which meant a managed install had no way to obtain it at
+all. The server then degrades SILENTLY to amplitude fallback (wearer vs Ext), so
+voice training appears to run and never learns anything. A beta user hit exactly
+that on 2026-08-25 and reported it as "voice training didn't work" -- it did run;
+it had nothing to train against.
+
+    npx --yes @gotcos/glasses-server@latest --setup-speaker-model
+
+- Downloads to ~/.cos-glasses/models/, which survives server updates. Anything
+  inside the package is destroyed by the next one.
+- Pinned SHA-256, verified BEFORE the file is moved into place. A mismatch is
+  deleted, not installed -- this file is handed to a native loader. The hash was
+  checked against the model on a working install and matches byte for byte.
+- Idempotent: an already-correct file is verified and left alone. A corrupted one
+  is replaced.
+- Synchronous, using the same curl idiom as the whisper download. The first
+  version was async and fire-and-forget, and because bin/cli.cjs is CJS with no
+  top-level await the command fell straight through into normal server startup
+  and installed nothing. Found by running it, not by reading it.
+
+The 26 MB stays out of the tarball; the install path is now a command instead of
+a paragraph.
+
+222 files, 3112 tests.
+
 ## 6.37.1
 
 Standalone meeting summaries default ON.

package/README.md

@@ -199,6 +199,16 @@ nothing; the COS companion sends its own wearer/`Ext` labels). Named
 per-speaker diarization needs a ~26 MB voiceprint model that is deliberately
 **not** shipped in the npm package, so it is a bolt-on:
 
+```bash
+npx --yes @gotcos/glasses-server@latest --setup-speaker-model
+```
+
+That downloads the model to `~/.cos-glasses/models/`, verifies it against a
+pinned SHA-256, and refuses to install anything that does not match. Restart the
+server afterwards and check `/api/health` — `speaker_id` should read `active`.
+
+To place it by hand instead:
+
 ```bash
 mkdir -p ~/.cos-glasses/models
 # put 3dspeaker_speech_eres2net_sv_en_voxceleb_16k.onnx there, then restart

package/bin/cli.cjs

@@ -26,6 +26,7 @@ const PKG_ROOT = resolve(__dirname, '..')
 const CONFIG_DIR = join(homedir(), '.cos-glasses')
 const PREPARE_ONLY = process.argv.includes('--prepare-only')
 const SETUP_TRANSCRIPTION = process.argv.includes('--setup-transcription')
+const SETUP_SPEAKER_MODEL = process.argv.includes('--setup-speaker-model')
 function optionValue(name) {
   const index = process.argv.indexOf(name)
   return index >= 0 ? process.argv[index + 1] : undefined
@@ -53,6 +54,104 @@ const yellow = (s) => `\x1b[33m${s}\x1b[0m`
 const dim = (s) => `\x1b[2m${s}\x1b[0m`
 const bold = (s) => `\x1b[1m${s}\x1b[0m`
 
+/**
+ * The ~26 MB voiceprint model, fetched on request.
+ *
+ * WHY THIS EXISTS. Named per-speaker diarization needs this model, and it is
+ * deliberately NOT in the npm tarball -- 26 MB on every install for a feature
+ * most users never turn on. The README documented a manual bolt-on ("put the
+ * .onnx there"), which left a managed install with no way to obtain it: the
+ * server then degrades silently to amplitude fallback (wearer vs Ext), so voice
+ * training appears to run and never learns anything. A beta user hit exactly
+ * that on 2026-08-25 and reported it as "voice training didn't work".
+ *
+ * Installed to ~/.cos-glasses/models/ deliberately. Anything inside the package
+ * is destroyed by the next update; the data home survives.
+ *
+ * The SHA-256 below was verified against the model on a working install and
+ * matched byte for byte. A download that does not match is DELETED, not
+ * installed -- this file is handed to a native loader.
+ */
+const SPEAKER_MODEL = {
+  filename: '3dspeaker_speech_eres2net_sv_en_voxceleb_16k.onnx',
+  url: 'https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/3dspeaker_speech_eres2net_sv_en_voxceleb_16k.onnx',
+  sha256: 'c59158379255ad66e161679cca6af8d52d51e389e3224ab7d7a7baae295c2db5',
+  bytes: 26485263,
+}
+
+function sha256File(path) {
+  const { createHash } = require('crypto')
+  return createHash('sha256').update(readFileSync(path)).digest('hex')
+}
+
+function setupSpeakerModel() {
+  const modelsDir = join(CONFIG_DIR, 'models')
+  const dest = join(modelsDir, SPEAKER_MODEL.filename)
+
+  if (existsSync(dest)) {
+    const have = sha256File(dest)
+    if (have === SPEAKER_MODEL.sha256) {
+      console.log('  Voiceprint model already installed and verified.')
+      console.log(`  ${dest}`)
+      return 0
+    }
+    console.log('  A file is present but does not match the expected checksum.')
+    console.log(`  Replacing it. (found ${have.slice(0, 16)}...)`)
+    try { unlinkSync(dest) } catch {}
+  }
+
+  mkdirSync(modelsDir, { recursive: true })
+  const partial = `${dest}.partial`
+  console.log('  Downloading the voiceprint model (~26 MB)...')
+
+  // SYNCHRONOUS on purpose. bin/cli.cjs is CJS with no top-level await, so an
+  // async download does not block the rest of this file -- the first version of
+  // this command fell straight through into normal server startup and installed
+  // nothing. Same curl idiom the whisper model download already uses.
+  try {
+    execFileSync('curl', [
+      '-fL',
+      '--retry', '5',
+      '--retry-all-errors',
+      '--retry-delay', '2',
+      '--connect-timeout', '30',
+      '--progress-bar',
+      SPEAKER_MODEL.url,
+      '-o', partial,
+    ], { stdio: 'inherit', timeout: 10 * 60_000 })
+  } catch (err) {
+    try { if (existsSync(partial)) unlinkSync(partial) } catch {}
+    console.error(`  Download failed: ${err && err.message ? err.message : err}`)
+    return 1
+  }
+
+  // VERIFY BEFORE INSTALLING. Never rename unverified bytes into the path a
+  // native model loader reads from.
+  const got = sha256File(partial)
+  if (got !== SPEAKER_MODEL.sha256) {
+    try { unlinkSync(partial) } catch {}
+    console.error('  Checksum mismatch - refusing to install.')
+    console.error(`    expected ${SPEAKER_MODEL.sha256}`)
+    console.error(`    got      ${got}`)
+    return 1
+  }
+
+  renameSync(partial, dest)
+  console.log('  Installed and verified:')
+  console.log(`  ${dest}`)
+  console.log('')
+  console.log('  Restart the server for it to take effect.')
+  console.log('  Then check /api/health -> speaker_id: it should read "active".')
+  return 0
+}
+
+if (SETUP_SPEAKER_MODEL) {
+  console.log('')
+  console.log(bold('  COS Glasses - voiceprint model'))
+  console.log('')
+  process.exit(setupSpeakerModel())
+}
+
 if (process.argv.includes('--help') || process.argv.includes('-h')) {
   console.log('')
   console.log(bold('  COS Glasses Server'))
@@ -61,6 +160,7 @@ if (process.argv.includes('--help') || process.argv.includes('-h')) {
   console.log('    npx --yes @gotcos/glasses-server@latest')
   console.log('    npx --yes @gotcos/glasses-server@latest --setup-transcription')
   console.log('    npx --yes @gotcos/glasses-server@latest --setup-transcription --transcription-tier balanced|max')
+  console.log('    npx --yes @gotcos/glasses-server@latest --setup-speaker-model')
   console.log('    npx --yes @gotcos/glasses-server@latest --prepare-only')
   console.log('')
   console.log('  Requirements:')

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.37.1",
+  "version": "6.37.2",
   "description": "COS Glasses — self-hosted AI heads-up-display server for Even G2 smart glasses, powered by Claude Code, Codex, or Cursor Agent CLI",
   "type": "module",
   "bin": {

package/server/routes/transcribe-stream.ts

@@ -61,14 +61,6 @@ import {
   appendChunkEmbedding,
   sweepExpiredChunkEmbeddings,
 } from '../lib/chunk-embedding-store.js'
-import {
-  evenSpeakerRoleMode,
-  formatEvenRoleAgreement,
-  parseEvenHubSpeakerRoleBody,
-  parseEvenHubSpeakerRoleQuery,
-  warnEvenSpeakerRoleApplyNotImplemented,
-  type EvenSpeakerRoleHistogram,
-} from '../lib/even-hub-speaker-role.js'
 import { archiveSessionAudio, runMeetingAudioRetention } from '../lib/meeting-audio-archive.js'
 import {
   countChunkWavs,
@@ -284,7 +276,6 @@ export interface TranscriptChunk {
   latencyMs?: number
   audioSha256?: string
   canonical?: boolean
-  evenHubSpeakerRole?: EvenSpeakerRoleHistogram
 }
 
 export interface ProviderCandidateRecord {
@@ -1954,11 +1945,8 @@ async function processStreamChunk(opts: {
   clientElapsed?: number
   /** Original client recording start, applied only before canonical chunks. */
   startTimeOverride?: number
-  evenHubSpeakerRole?: EvenSpeakerRoleHistogram
 }): Promise<StreamChunkCompletionResponse> {
   const { sessionId, chunkIndex, clientSpeaker, audioBuffer, candidate } = opts
-  const evenHubSpeakerRole = evenSpeakerRoleMode() === 'off' ? undefined : opts.evenHubSpeakerRole
-  if (evenHubSpeakerRole) warnEvenSpeakerRoleApplyNotImplemented()
   const tReq = performance.now()
   validateSessionId(sessionId)
   validateChunkIndex(chunkIndex)
@@ -2078,15 +2066,6 @@ async function processStreamChunk(opts: {
   }
 
   const { speaker, similarity } = await speakerPromise
-  if (evenHubSpeakerRole) {
-    console.log(formatEvenRoleAgreement({
-      chunkIndex,
-      even: evenHubSpeakerRole,
-      amp: clientSpeaker,
-      emb: speaker,
-      similarity,
-    }))
-  }
   // Client time is authoritative for live network jitter and deferred replay.
   const elapsed = Number.isFinite(opts.clientElapsed) && (opts.clientElapsed as number) >= 0
     ? Math.round(opts.clientElapsed as number)
@@ -2131,7 +2110,6 @@ async function processStreamChunk(opts: {
     latencyMs,
     audioSha256,
     canonical: true,
-    evenHubSpeakerRole,
   }
   const finalExisting = session.chunks[chunkIndex]
   if (finalExisting?.text) {
@@ -2285,7 +2263,6 @@ transcribeStreamRouter.post('/transcribe-stream', async (req, res) => {
       audioBuffer,
       clientElapsed,
       startTimeOverride,
-      evenHubSpeakerRole: parseEvenHubSpeakerRoleQuery(req.query.eh),
     }))
   } catch (err: unknown) {
     sendStreamError(res, err)
@@ -2350,7 +2327,6 @@ transcribeStreamRouter.post('/transcribe-stream/offline-sessions/:sessionId/chun
       clientSpeaker: String(body.clientSpeaker ?? 'Unknown'),
       audioBuffer,
       clientElapsed: typeof body.elapsed === 'number' ? body.elapsed : Number(body.elapsed ?? 0),
-      evenHubSpeakerRole: parseEvenHubSpeakerRoleBody(body.evenHubSpeakerRole),
       candidate: {
         provider: 'iphone-whisperkit-beta',
         text: normalizeCandidateText(candidate.text),
@@ -2429,7 +2405,6 @@ transcribeStreamRouter.post('/transcribe-stream/candidates', async (req, res) =>
       clientSpeaker: String(body.clientSpeaker ?? 'Unknown'),
       audioBuffer,
       clientElapsed: typeof body.elapsed === 'number' ? body.elapsed : Number(body.elapsed ?? 0),
-      evenHubSpeakerRole: parseEvenHubSpeakerRoleBody(body.evenHubSpeakerRole),
       candidate: {
         provider: 'iphone-whisperkit-beta',
         text: normalizeCandidateText(candidate.text),

package/server/lib/even-hub-speaker-role.ts

@@ -1,116 +0,0 @@
-// Even Hub 0.0.14 wearer-vs-other histogram, carried on a meeting chunk.
-// Identity is a suggestion. This module parses and logs. It does not name
-// people and does not change identifyChunkSpeaker.
-
-export type EvenSpeakerRole = 'self' | 'other' | 'unknown'
-export type EvenSpeakerRoleMajority = EvenSpeakerRole | 'tie'
-
-export interface EvenSpeakerRoleHistogram {
-  schema: 1
-  frames: number
-  self: number
-  other: number
-  unknown: number
-  majority: EvenSpeakerRoleMajority
-  directionPresent: number
-  directionLast: number | null
-}
-
-export type EvenSpeakerRoleMode = 'off' | 'log' | 'apply'
-
-export function evenSpeakerRoleMode(): EvenSpeakerRoleMode {
-  const raw = (process.env.COS_EVEN_SPEAKER_ROLE ?? 'log').trim().toLowerCase()
-  if (raw === 'off' || raw === '0' || raw === 'false') return 'off'
-  if (raw === 'apply') return 'apply'
-  return 'log'
-}
-
-let applyNotImplementedWarned = false
-
-/** Gate A is not in this slice. apply must not silently change labels. */
-export function warnEvenSpeakerRoleApplyNotImplemented(): void {
-  if (evenSpeakerRoleMode() !== 'apply' || applyNotImplementedWarned) return
-  applyNotImplementedWarned = true
-  console.warn('[even-role] COS_EVEN_SPEAKER_ROLE=apply is not implemented; logging only')
-}
-
-function majorityOf(self: number, other: number, unknown: number, frames: number): EvenSpeakerRoleMajority {
-  if (frames <= 0) return 'unknown'
-  if (self > other && self > unknown) return 'self'
-  if (other > self && other > unknown) return 'other'
-  if (unknown > self && unknown > other) return 'unknown'
-  return 'tie'
-}
-
-function asNonNegInt(raw: unknown): number | null {
-  const n = typeof raw === 'number' ? raw : typeof raw === 'string' && raw !== '' ? Number(raw) : NaN
-  if (!Number.isFinite(n) || n < 0 || !Number.isInteger(n)) return null
-  return n
-}
-
-/** Compact query `eh=self,other,unknown,frames,directionPresent,directionLast`. */
-export function parseEvenHubSpeakerRoleQuery(raw: unknown): EvenSpeakerRoleHistogram | undefined {
-  if (typeof raw !== 'string' || raw.length === 0) return undefined
-  const parts = raw.split(',')
-  if (parts.length < 4 || parts.length > 6) return undefined
-  const self = asNonNegInt(parts[0])
-  const other = asNonNegInt(parts[1])
-  const unknown = asNonNegInt(parts[2])
-  const frames = asNonNegInt(parts[3])
-  if (self == null || other == null || unknown == null || frames == null) return undefined
-  if (self + other + unknown !== frames) return undefined
-  const directionPresent = parts.length >= 5 ? asNonNegInt(parts[4]) : 0
-  if (directionPresent == null) return undefined
-  let directionLast: number | null = null
-  if (parts.length === 6 && parts[5] !== '') {
-    const last = Number(parts[5])
-    if (!Number.isFinite(last)) return undefined
-    directionLast = last
-  }
-  return {
-    schema: 1,
-    frames,
-    self,
-    other,
-    unknown,
-    majority: majorityOf(self, other, unknown, frames),
-    directionPresent,
-    directionLast,
-  }
-}
-
-export function parseEvenHubSpeakerRoleBody(raw: unknown): EvenSpeakerRoleHistogram | undefined {
-  if (!raw || typeof raw !== 'object') return undefined
-  const o = raw as Record<string, unknown>
-  const self = asNonNegInt(o.self)
-  const other = asNonNegInt(o.other)
-  const unknown = asNonNegInt(o.unknown)
-  const frames = asNonNegInt(o.frames)
-  if (self == null || other == null || unknown == null || frames == null) return undefined
-  if (self + other + unknown !== frames) return undefined
-  const directionPresent = o.directionPresent == null ? 0 : asNonNegInt(o.directionPresent)
-  if (directionPresent == null) return undefined
-  const directionLast = o.directionLast == null || o.directionLast === ''
-    ? null
-    : (typeof o.directionLast === 'number' && Number.isFinite(o.directionLast) ? o.directionLast : null)
-  return {
-    schema: 1,
-    frames,
-    self,
-    other,
-    unknown,
-    majority: majorityOf(self, other, unknown, frames),
-    directionPresent,
-    directionLast,
-  }
-}
-
-export function formatEvenRoleAgreement(opts: {
-  chunkIndex: number
-  even: EvenSpeakerRoleHistogram
-  amp: string
-  emb: string
-  similarity: number
-}): string {
-  return `[even-role] chunk=${opts.chunkIndex} even=${opts.even.majority} amp=${opts.amp} emb=${opts.emb} sim=${opts.similarity.toFixed(2)} frames=${opts.even.frames}`
-}