feat(tts): curated Kokoro+Piper radio voices, classroom instructor from pool

Controllers now speak with the four Kokoro speakers (bm_george, bf_emma,
am_michael, af_heart — heard most, best quality), pilots with six distinct
Piper speakers (alan, alba, joe, amy, bryce, kristin); ATIS moves to the
dedicated ljspeech broadcast voice. The classroom instructor draws a stable
controller-pool voice per module when the user hasn't configured one, and the
TTS cache key follows the effective voice.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

app/pages/classroom.vue

@@ -1537,6 +1537,7 @@ import {looksLikeCallsignKey, matchTranscriptionToFields, type SttFieldDef} from
 import {loadPizzicatoLite} from '~~/shared/utils/pizzicatoLite'
 import type {PizzicatoLite} from '~~/shared/utils/pizzicatoLite'
 import {createNoiseGenerators, getReadabilityProfile} from '~~/shared/utils/radioEffects'
+import {controllerVoiceFor} from '~~/shared/utils/voicePool'
 import {DEFAULT_AIRLINE_TELEPHONY, normalizeRadioPhrase, normalizeMetarPhrase} from '~~/shared/utils/radioSpeech'
 
 definePageMeta({middleware: ['require-auth', 'require-classroom-intro']})
@@ -4649,8 +4650,19 @@ async function ensurePizzicato(ctx: AudioContext | null): Promise<PizzicatoLite
   return pizzicatoLiteInstance
 }
 
+/**
+ * The instructor's voice: the user's explicit choice, or a stable
+ * controller-pool voice per module — each module sounds like its own
+ * instructor, consistent across the whole lesson.
+ */
+function instructorVoice(): string {
+  const configured = cfg.value.voice?.trim()
+  if (configured) return configured
+  return controllerVoiceFor(`classroom:${current.value?.id || 'learn'}`)
+}
+
 function buildSayCacheKey(text: string, rate: number): string {
-  const voice = cfg.value.voice?.trim().toLowerCase() || 'default'
+  const voice = instructorVoice().toLowerCase()
   const radioLevel = cfg.value.radioLevel
   return `${voice}|${radioLevel}|${rate.toFixed(2)}|${text}`
 }
@@ -5007,9 +5019,7 @@ async function say(text: string) {
     tag: 'learn-target'
   }
 
-  if (cfg.value.voice) {
-    payload.voice = cfg.value.voice
-  }
+  payload.voice = instructorVoice()
 
   const cacheKey = buildSayCacheKey(speakText, normalizedRate)
 

server/utils/voiceRegistry.test.ts

@@ -1,26 +1,32 @@
 import { describe, it } from 'node:test'
 import assert from 'node:assert/strict'
-import { resolveSpeachesVoice, SPEACHES_VOICE_MAP } from '~~/server/utils/voiceRegistry'
+import {
+  ATIS_SPEACHES_VOICE,
+  KOKORO_MODEL,
+  resolveSpeachesVoice,
+  SPEACHES_VOICE_MAP,
+} from '~~/server/utils/voiceRegistry'
 
 const fallback = { model: 'speaches-ai/piper-de_DE-thorsten-medium', voice: 'de_DE-thorsten-medium' }
 
 describe('resolveSpeachesVoice', () => {
-  it('maps every pool voice to a matching piper model/voice pair', () => {
+  it('maps every pool voice to a consistent model/voice pair', () => {
     for (const [logical, mapped] of Object.entries(SPEACHES_VOICE_MAP)) {
       const result = resolveSpeachesVoice(logical, fallback)
-      assert.equal(result.voice, mapped.voice)
-      assert.equal(result.model, `speaches-ai/piper-${mapped.voice}`)
+      assert.deepEqual(result, mapped)
+      if (result.model === KOKORO_MODEL) {
+        // Kokoro voice ids look like af_heart / bm_george.
+        assert.match(result.voice, /^[abehijpz][fm]_[a-z_]+$/)
+      } else {
+        assert.equal(result.model, `speaches-ai/piper-${result.voice}`)
+        assert.match(result.voice, /^en_/)
+      }
     }
   })
 
-  it('maps the default controller voice to an English speaker', () => {
-    const result = resolveSpeachesVoice('alloy', fallback)
-    assert.match(result.voice, /^en_/)
-  })
-
   it('is case-insensitive and trims', () => {
     const result = resolveSpeachesVoice('  Alloy ', fallback)
-    assert.equal(result.voice, SPEACHES_VOICE_MAP.alloy!.voice)
+    assert.deepEqual(result, SPEACHES_VOICE_MAP.alloy)
   })
 
   it('falls back to the configured pair for unknown ids', () => {
@@ -30,8 +36,7 @@ describe('resolveSpeachesVoice', () => {
 
   it('gives the ATIS broadcast its own dedicated speaker', () => {
     const atis = resolveSpeachesVoice('verse', fallback, 'atis')
-    assert.match(atis.voice, /^en_/)
-    assert.notEqual(atis.voice, resolveSpeachesVoice('verse', fallback).voice)
+    assert.deepEqual(atis, ATIS_SPEACHES_VOICE)
     const allPoolVoices = Object.values(SPEACHES_VOICE_MAP).map(v => v.voice)
     assert.equal(allPoolVoices.includes(atis.voice), false, 'ATIS voice must not be in any pool')
   })
@@ -44,4 +49,9 @@ describe('resolveSpeachesVoice', () => {
       assert.equal(pilots.includes(voice), false, `${voice} used for both controller and pilot`)
     }
   })
+
+  it('assigns no speaker twice across the whole registry', () => {
+    const voices = [...Object.values(SPEACHES_VOICE_MAP).map(v => v.voice), ATIS_SPEACHES_VOICE.voice]
+    assert.equal(new Set(voices).size, voices.length, `duplicate speaker in ${voices.join(', ')}`)
+  })
 })

server/utils/voiceRegistry.ts

@@ -1,36 +1,42 @@
 /**
  * Maps the product-wide logical voice ids (shared/utils/voicePool.ts — OpenAI
- * voice names) to Speaches/Piper model+voice pairs. The client keeps sending
- * pool names; only the Speaches branch of /api/atc/say resolves them here, so
- * the same ids work unchanged on the OpenAI provider.
+ * voice names) to Speaches model+voice pairs. The client keeps sending pool
+ * names; only the Speaches branch of /api/atc/say resolves them here, so the
+ * same ids work unchanged on the OpenAI provider.
  *
- * Speaches auto-downloads missing `speaches-ai/piper-*` models on first use.
- * Controller voices (alloy/echo/onyx/sage) and pilot voices (verse + the
- * PILOT_VOICES pool) map to disjoint speakers so traffic never sounds like
- * the controller talking to itself.
+ * Speaches auto-downloads missing `speaches-ai/piper-*` models on first use;
+ * the Kokoro voices all live in one multi-voice model. Controllers get the
+ * Kokoro speakers (heard most, best quality), pilots the distinct Piper
+ * speakers — the pools stay disjoint so traffic never sounds like the
+ * controller talking to itself. Selection follows the curated radio set
+ * (see docs/plans/2026-07-19-phraseology-taxi-finetune-design.md).
  */
 
 export type SpeachesVoice = { model: string; voice: string }
 
+export const KOKORO_MODEL = 'speaches-ai/Kokoro-82M-v1.0-ONNX'
+
 const piper = (voice: string): SpeachesVoice => ({
   model: `speaches-ai/piper-${voice}`,
   voice,
 })
 
+const kokoro = (voice: string): SpeachesVoice => ({ model: KOKORO_MODEL, voice })
+
 export const SPEACHES_VOICE_MAP: Record<string, SpeachesVoice> = {
   // Controller pool (CONTROLLER_VOICES)
-  alloy: piper('en_US-ryan-medium'),
-  echo: piper('en_GB-alan-medium'),
-  onyx: piper('en_US-john-medium'),
-  sage: piper('en_GB-jenny_dioco-medium'),
+  alloy: kokoro('bm_george'),
+  echo: kokoro('bf_emma'),
+  onyx: kokoro('am_michael'),
+  sage: kokoro('af_heart'),
   // The user's own readback voice (speakPilotReadback)
   verse: piper('en_US-lessac-medium'),
   // Simulated pilot pool (PILOT_VOICES)
-  ash: piper('en_US-joe-medium'),
-  ballad: piper('en_GB-northern_english_male-medium'),
-  coral: piper('en_US-amy-medium'),
-  fable: piper('en_US-danny-low'),
-  nova: piper('en_US-hfc_female-medium'),
+  ash: piper('en_GB-alan-medium'),
+  ballad: piper('en_GB-alba-medium'),
+  coral: piper('en_US-joe-medium'),
+  fable: piper('en_US-amy-medium'),
+  nova: piper('en_US-bryce-medium'),
   shimmer: piper('en_US-kristin-medium'),
 }
 
@@ -38,7 +44,7 @@ export const SPEACHES_VOICE_MAP: Record<string, SpeachesVoice> = {
  * ATIS is a robotic broadcast in the real world — it gets a dedicated speaker
  * outside every pool so it never sounds like a controller or a pilot.
  */
-export const ATIS_SPEACHES_VOICE: SpeachesVoice = piper('en_GB-alba-medium')
+export const ATIS_SPEACHES_VOICE: SpeachesVoice = piper('en_US-ljspeech-high')
 
 export function resolveSpeachesVoice(
   logical: string,