@qvac/tts-ggml 0.9.2 → 0.10.1

package/test/integration/multiple-runs.test.js

@@ -29,13 +29,16 @@ const {
 } = require('../utils/runChatterboxTTS')
 const { loadSupertonicTTS, runSupertonicTTS } = require('../utils/runSupertonicTTS')
 const { ensureChatterboxModels, ensureSupertonicModel } = require('../utils/downloadModel')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { NO_GPU, GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
 
-// Lifecycle / sequential-run test, not a GPU policy test: rely on the
-// package default (`useGPU: false`) rather than opting into GPU here.
-// Tests that *are* about GPU live in gpu-smoke.test.js.
+// Lifecycle / sequential-run tests, not GPU policy tests. The Chatterbox ones
+// run several syntheses, so they run on the GPU only and are skipped on CPU-only
+// runners (NO_GPU=true); the Supertonic reload keeps the package default
+// (`useGPU: false`). Tests that *are* about GPU live in gpu-smoke.test.js.
 
 function getBaseDir() {
   return isMobile && global.testDir ? global.testDir : '.'
@@ -49,7 +52,7 @@ const PHRASES = [
 
 test(
   'Chatterbox: multiple sequential runs reuse the same engine instance',
-  { timeout: 1800000 },
+  { timeout: 1800000, skip: NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureChatterboxModels({ targetDir: path.join(baseDir, 'models') })
@@ -73,9 +76,11 @@ test(
     }
 
     const model = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
       modelDir: download.targetDir,
       refWavPath,
-      language: 'en'
+      language: 'en',
+      useGPU: GPU_ONLY_USE_GPU
     })
     try {
       const timings = []
@@ -109,7 +114,7 @@ test(
 
 test(
   'Chatterbox: fresh instance per run (app-restart simulation)',
-  { timeout: 1800000 },
+  { timeout: 1800000, skip: NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureChatterboxModels({ targetDir: path.join(baseDir, 'models') })
@@ -137,9 +142,11 @@ test(
     for (let i = 0; i < N; i++) {
       const t0 = Date.now()
       const model = await loadChatterboxTTS({
+        threads: TTS_TEST_THREADS,
         modelDir: download.targetDir,
         refWavPath,
-        language: 'en'
+        language: 'en',
+        useGPU: GPU_ONLY_USE_GPU
       })
       const loadMs = Date.now() - t0
       try {
@@ -172,57 +179,63 @@ test(
   }
 )
 
-test('Chatterbox: reload() between runs preserves stability', { timeout: 1800000 }, async (t) => {
-  const baseDir = getBaseDir()
-  const download = await ensureChatterboxModels({ targetDir: path.join(baseDir, 'models') })
-  if (!download.success) {
-    t.fail(
-      'Chatterbox GGUFs not available - registry fetch failed. Run `npm run download-models:registry` or stage models locally.'
-    )
-    return
-  }
+test(
+  'Chatterbox: reload() between runs preserves stability',
+  { timeout: 1800000, skip: NO_GPU },
+  async (t) => {
+    const baseDir = getBaseDir()
+    const download = await ensureChatterboxModels({ targetDir: path.join(baseDir, 'models') })
+    if (!download.success) {
+      t.fail(
+        'Chatterbox GGUFs not available - registry fetch failed. Run `npm run download-models:registry` or stage models locally.'
+      )
+      return
+    }
 
-  // Mobile-aware resolution: on iOS / Android the asset is staged into
-  // `Library/Caches/jfk.wav` via `global.assetPaths`; on desktop falls
-  // back to the in-tree `test/reference-audio/jfk.wav`. The bundled
-  // path is not readable from native code on iOS, which would trip
-  // `ModelFileNotFound` the moment `model.load()` reaches the C++
-  // `ChatterboxModel::validateConfig` check.
-  const refWavPath = resolveRefWavPath({})
-  if (!fs.existsSync(refWavPath)) {
-    t.pass('Skipped: reference audio missing')
-    return
-  }
+    // Mobile-aware resolution: on iOS / Android the asset is staged into
+    // `Library/Caches/jfk.wav` via `global.assetPaths`; on desktop falls
+    // back to the in-tree `test/reference-audio/jfk.wav`. The bundled
+    // path is not readable from native code on iOS, which would trip
+    // `ModelFileNotFound` the moment `model.load()` reaches the C++
+    // `ChatterboxModel::validateConfig` check.
+    const refWavPath = resolveRefWavPath({})
+    if (!fs.existsSync(refWavPath)) {
+      t.pass('Skipped: reference audio missing')
+      return
+    }
 
-  const model = await loadChatterboxTTS({
-    modelDir: download.targetDir,
-    refWavPath,
-    language: 'en'
-  })
-  try {
-    const r1 = await runChatterboxTTS(
-      model,
-      { text: 'First run before reload.' },
-      { minSamples: 5000 }
-    )
-    t.ok(r1.passed, 'first run before reload should pass')
+    const model = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
+      modelDir: download.targetDir,
+      refWavPath,
+      language: 'en',
+      useGPU: GPU_ONLY_USE_GPU
+    })
+    try {
+      const r1 = await runChatterboxTTS(
+        model,
+        { text: 'First run before reload.' },
+        { minSamples: 5000 }
+      )
+      t.ok(r1.passed, 'first run before reload should pass')
 
-    await model.reload({ language: 'en' })
-    t.pass('reload() resolved')
+      await model.reload({ language: 'en' })
+      t.pass('reload() resolved')
 
-    const r2 = await runChatterboxTTS(
-      model,
-      { text: 'Second run after reload.' },
-      { minSamples: 5000 }
-    )
-    t.ok(r2.passed, 'second run after reload should pass')
-    t.ok(r2.data.sampleCount > 0, 'reloaded model produces audio')
-  } finally {
-    try {
-      await model.unload()
-    } catch (_e) {}
+      const r2 = await runChatterboxTTS(
+        model,
+        { text: 'Second run after reload.' },
+        { minSamples: 5000 }
+      )
+      t.ok(r2.passed, 'second run after reload should pass')
+      t.ok(r2.data.sampleCount > 0, 'reloaded model produces audio')
+    } finally {
+      try {
+        await model.unload()
+      } catch (_e) {}
+    }
   }
-})
+)
 
 test('Supertonic: reload() between runs preserves stability', { timeout: 1800000 }, async (t) => {
   const baseDir = getBaseDir()
@@ -235,6 +248,7 @@ test('Supertonic: reload() between runs preserves stability', { timeout: 1800000
   }
 
   const model = await loadSupertonicTTS({
+    threads: TTS_TEST_THREADS,
     supertonicModelPath: download.path,
     voice: 'F1',
     language: 'en',
@@ -267,7 +281,7 @@ test('Supertonic: reload() between runs preserves stability', { timeout: 1800000
 
 test(
   'Engine swap: chatterbox -> supertonic -> chatterbox in separate instances',
-  { timeout: 1800000 },
+  { timeout: 1800000, skip: NO_GPU },
   async (t) => {
     const baseDir = getBaseDir()
     const cb = await ensureChatterboxModels({ targetDir: path.join(baseDir, 'models') })
@@ -291,7 +305,13 @@ test(
       return
     }
 
-    const c1 = await loadChatterboxTTS({ modelDir: cb.targetDir, refWavPath, language: 'en' })
+    const c1 = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
+      modelDir: cb.targetDir,
+      refWavPath,
+      language: 'en',
+      useGPU: GPU_ONLY_USE_GPU
+    })
     try {
       const r = await runChatterboxTTS(c1, { text: 'Hello from chatterbox.' }, { minSamples: 5000 })
       t.ok(r.passed, 'first chatterbox instance OK')
@@ -302,10 +322,11 @@ test(
     }
 
     const s1 = await loadSupertonicTTS({
+      threads: TTS_TEST_THREADS,
       supertonicModelPath: st.path,
       voice: 'F1',
       language: 'en',
-      useGPU: false
+      useGPU: GPU_ONLY_USE_GPU
     })
     try {
       const r = await runSupertonicTTS(s1, { text: 'Hello from supertonic.' }, { minSamples: 5000 })
@@ -316,7 +337,13 @@ test(
       } catch (_e) {}
     }
 
-    const c2 = await loadChatterboxTTS({ modelDir: cb.targetDir, refWavPath, language: 'en' })
+    const c2 = await loadChatterboxTTS({
+      threads: TTS_TEST_THREADS,
+      modelDir: cb.targetDir,
+      refWavPath,
+      language: 'en',
+      useGPU: GPU_ONLY_USE_GPU
+    })
     try {
       const r = await runChatterboxTTS(
         c2,

package/test/integration/output-sample-rate.test.js

@@ -12,6 +12,8 @@ const test = require('brittle')
 const TTSGgml = require('@qvac/tts-ggml')
 
 const { ensureSupertonicModel, ensureParlerModel } = require('../utils/downloadModel')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { SKIP_PARLER_GPU_ONLY, PARLER_GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -40,6 +42,7 @@ async function collect(model, text) {
 // JS half, which must forward the combination rather than reject it up front.
 test('CosyVoice3: enhancer + streaming forwards a non-native outputSampleRate', (t) => {
   const model = new TTSGgml({
+    threads: TTS_TEST_THREADS,
     engine: TTSGgml.ENGINE_COSYVOICE3,
     files: {
       cosyvoiceModelDir: './models/cosyvoice3',
@@ -72,6 +75,7 @@ test(
     const text = 'Output rate selection resamples the synthesized audio.'
 
     const native = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_SUPERTONIC,
       files: { supertonicModel: dl.path },
       voice: 'F1',
@@ -91,6 +95,7 @@ test(
     }
 
     const resampled = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_SUPERTONIC,
       files: { supertonicModel: dl.path },
       voice: 'F1',
@@ -119,7 +124,7 @@ test(
 
 test(
   'Parler: outputSampleRate=16000 resamples and reports 16 kHz',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const dl = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -131,10 +136,12 @@ test(
     const text = 'Output rate selection resamples the synthesized audio.'
 
     const native = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: dl.path },
       voice: 'Laura',
       seed: 42,
+      config: { useGPU: PARLER_GPU_ONLY_USE_GPU },
       opts: { stats: true }
     })
     await native.load()
@@ -150,11 +157,12 @@ test(
     }
 
     const resampled = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: dl.path },
       voice: 'Laura',
       seed: 42,
-      config: { outputSampleRate: 16000 },
+      config: { outputSampleRate: 16000, useGPU: PARLER_GPU_ONLY_USE_GPU },
       opts: { stats: true }
     })
     await resampled.load()

package/test/integration/parler-wer.test.js

@@ -7,9 +7,9 @@
 // chatterbox WER leg in addon.test.js (whisper ASR + wordErrorRate).
 //
 // macOS-desktop only (isDarwin): the whisper WER infra is desktop-only
-// across the tts-ggml suite. Runs on both the Metal GPU runner and the
-// no_gpu CPU runner (backend chosen like the quant loop). large-v1 is
-// RAM-gated; both tiers skip cleanly if their GGUF isn't staged.
+// across the tts-ggml suite. Runs on the Metal GPU runner only and is skipped
+// on CPU-only runners (NO_GPU=true). large-v1 is RAM-gated; both tiers skip
+// cleanly if their GGUF isn't staged.
 
 const os = require('bare-os')
 const path = require('bare-path')
@@ -19,6 +19,7 @@ const test = require('brittle')
 const { loadParlerTTS, runParlerTTS } = require('../utils/runParlerTTS')
 const { ensureParlerModel, ensureWhisperModel } = require('../utils/downloadModel')
 const { loadWhisper, runWhisper } = require('../utils/runWhisper')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
 
 const platform = os.platform()
 const isDarwin = platform === 'darwin'
@@ -26,10 +27,9 @@ const isApple = platform === 'darwin' || platform === 'ios'
 const NO_GPU = proc.env && proc.env.NO_GPU === 'true'
 const LARGE_MIN_RAM_BYTES = 16 * 1024 ** 3
 
-// Calibrated 2026-07-24: the CPU (no_gpu) runner is 0.0% WER at any seed; the
-// Metal/GPU runner diverges (FP reorder) — seed 42 is 0% on Apple-Silicon but
-// 22.2% on Intel-Mac Metal — so 0.35 tolerates the GPU variance while the CPU
-// leg stays effectively strict (greedy would be 50-100%, so this still asserts it).
+// Calibrated 2026-07-24: Metal diverges by FP reorder — seed 42 is 0% on
+// Apple-Silicon but 22.2% on Intel-Mac Metal — so 0.35 tolerates the GPU
+// variance while still asserting the words (greedy would be 50-100%).
 const WER_THRESHOLD = 0.35
 
 // Clear English sentences, all >= 9 words (a single ASR slip stays <= ~0.11,
@@ -50,7 +50,7 @@ function getBaseDir() {
 for (const { variant } of PARLER_WER_TIERS) {
   test(
     `Parler WER (${variant}): English synthesis transcribes within ${(WER_THRESHOLD * 100).toFixed(0)}%`,
-    { timeout: 1800000, skip: !isDarwin },
+    { timeout: 1800000, skip: !isDarwin || NO_GPU },
     async (t) => {
       if (variant === 'large' && os.totalmem() < LARGE_MIN_RAM_BYTES) {
         t.pass(
@@ -77,7 +77,12 @@ for (const { variant } of PARLER_WER_TIERS) {
       }
 
       const useGPU = isApple && !NO_GPU
-      const model = await loadParlerTTS({ parlerModelPath: download.path, seed: 42, useGPU })
+      const model = await loadParlerTTS({
+        threads: TTS_TEST_THREADS,
+        parlerModelPath: download.path,
+        seed: 42,
+        useGPU
+      })
       const entries = []
       try {
         for (const text of WER_SENTENCES) {

package/test/integration/parler.test.js

@@ -7,17 +7,17 @@
 
 const os = require('bare-os')
 const path = require('bare-path')
-const proc = require('bare-process')
 const test = require('brittle')
 
 const { loadParlerTTS, runParlerTTS } = require('../utils/runParlerTTS')
 const { ensureParlerModel } = require('../utils/downloadModel')
 const { recordTtsStats } = require('../utils/perf-helper')
+const { TTS_TEST_THREADS } = require('../utils/testThreads')
+const { NO_GPU, SKIP_PARLER_GPU_ONLY, PARLER_GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
 const isApple = platform === 'darwin' || platform === 'ios'
-const NO_GPU = proc.env && proc.env.NO_GPU === 'true'
 
 function getBaseDir() {
   return isMobile && global.testDir ? global.testDir : '.'
@@ -39,7 +39,11 @@ test(
 
     // Deliberately NO description/voice/emotion: the engine renders the
     // models' recommended fallback caption — everything works on defaults.
-    const model = await loadParlerTTS({ parlerModelPath: download.path, seed: 42 })
+    const model = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
+      parlerModelPath: download.path,
+      seed: 42
+    })
     try {
       const wavPath = isMobile ? undefined : path.join(baseDir, 'test', 'output', 'parler-en.wav')
       const text = 'The parler engine speaks with a voice controlled by a text description.'
@@ -100,7 +104,11 @@ test(
       return
     }
 
-    const model = await loadParlerTTS({ parlerModelPath: download.path, voice: 'Laura' })
+    const model = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
+      parlerModelPath: download.path,
+      voice: 'Laura'
+    })
     try {
       const response = await model.run({
         type: 'text',
@@ -128,7 +136,7 @@ test(
 
 test(
   'Parler TTS (ggml): description matrix — conflicts throw, per-call wins, defaults work',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -143,6 +151,7 @@ test(
     t.exception(
       () =>
         new TTSGgml({
+          threads: TTS_TEST_THREADS,
           engine: TTSGgml.ENGINE_PARLER,
           files: { parlerModel: download.path },
           description: 'A calm female voice.',
@@ -157,6 +166,7 @@ test(
     t.exception(
       () =>
         new TTSGgml({
+          threads: TTS_TEST_THREADS,
           engine: TTSGgml.ENGINE_SUPERTONIC,
           files: { supertonicModel: download.path },
           emotion: 'happy'
@@ -167,6 +177,7 @@ test(
     t.exception(
       () =>
         new TTSGgml({
+          threads: TTS_TEST_THREADS,
           engine: TTSGgml.ENGINE_SUPERTONIC,
           files: { supertonicModel: download.path },
           pitch: 'high'
@@ -176,9 +187,11 @@ test(
     )
 
     const model = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       voice: 'Laura',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       // Per-call description conflict (same level) rejects.
@@ -218,9 +231,11 @@ test(
 
     // Constructor-level free-text description + per-call template rejects.
     const descModel = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       description: 'A calm female voice, very clear audio.',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       await t.exception(
@@ -238,7 +253,7 @@ test(
 
 test(
   'Parler TTS (ggml): emotion flag changes the audio; per-call switch needs no reload',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -248,9 +263,11 @@ test(
     }
 
     const model = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       voice: 'Laura',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       const text = 'The weather is wonderful today.'
@@ -285,7 +302,7 @@ test(
 
 test(
   'Parler TTS (ggml): fixed seed is deterministic across runs and reload',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -295,10 +312,12 @@ test(
     }
 
     const model = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       voice: 'Laura',
       emotion: 'neutral',
-      seed: 42
+      seed: 42,
+      useGPU: PARLER_GPU_ONLY_USE_GPU
     })
     try {
       const text = 'Deterministic synthesis check.'
@@ -329,7 +348,7 @@ test(
 
 test(
   'Parler TTS (ggml): runStream emits per-sentence chunks with one pinned description',
-  { timeout: 600000 },
+  { timeout: 600000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -340,9 +359,11 @@ test(
 
     const TTSGgml = require('@qvac/tts-ggml')
     const model = new TTSGgml({
+      threads: TTS_TEST_THREADS,
       engine: TTSGgml.ENGINE_PARLER,
       files: { parlerModel: download.path },
       seed: 42,
+      config: { useGPU: PARLER_GPU_ONLY_USE_GPU },
       opts: { stats: true }
     })
     await model.load()
@@ -376,7 +397,7 @@ test(
 
 test(
   'Parler TTS (ggml): native streaming emits monotonic chunks equal to batch',
-  { timeout: 900000 },
+  { timeout: 900000, skip: SKIP_PARLER_GPU_ONLY },
   async (t) => {
     const baseDir = getBaseDir()
     const download = await ensureParlerModel({ targetDir: path.join(baseDir, 'models') })
@@ -392,6 +413,7 @@ test(
 
     async function synth(extra) {
       const model = new TTSGgml({
+        threads: TTS_TEST_THREADS,
         engine: TTSGgml.ENGINE_PARLER,
         files: { parlerModel: download.path },
         seed: 42,
@@ -481,6 +503,7 @@ test(
     }
 
     const model = await loadParlerTTS({
+      threads: TTS_TEST_THREADS,
       parlerModelPath: download.path,
       voice: 'Rohit',
       emotion: 'happy',
@@ -507,22 +530,24 @@ test(
 
 // Desktop quant + backend coverage: q8_0 + f16 for indic/mini, q8_0 for large.
 // Runs on the platform's GPU-if-available backend (Metal on Apple GPU runners,
-// CPU otherwise). large self-gates on RAM; any tier that isn't staged skips
-// (t.pass) so a not-yet-registered tier never fails CI. Registered per PR #3372.
-// Desktop-only: the mobile Parler surface is covered by the gpu-smoke legs.
+// CPU otherwise). mini f16 and large are the slowest tiers on CPU (roughly
+// 45-80 s per synthesis), so they run on Metal only. large self-gates on RAM; any tier that isn't
+// staged skips (t.pass) so a not-yet-registered tier never fails CI. Registered
+// per PR #3372. Desktop-only: the mobile Parler surface is covered by the
+// gpu-smoke legs.
 const PARLER_QUANT_MATRIX = [
-  { variant: 'indic', quant: 'q8_0' },
-  { variant: 'indic', quant: 'f16' },
-  { variant: 'mini', quant: 'q8_0' },
-  { variant: 'mini', quant: 'f16' },
-  { variant: 'large', quant: 'q8_0' }
+  { variant: 'indic', quant: 'q8_0', gpuOnly: false },
+  { variant: 'indic', quant: 'f16', gpuOnly: false },
+  { variant: 'mini', quant: 'q8_0', gpuOnly: false },
+  { variant: 'mini', quant: 'f16', gpuOnly: true },
+  { variant: 'large', quant: 'q8_0', gpuOnly: true }
 ]
 const LARGE_MIN_RAM_BYTES = 16 * 1024 ** 3
 
-for (const { variant, quant } of PARLER_QUANT_MATRIX) {
+for (const { variant, quant, gpuOnly } of PARLER_QUANT_MATRIX) {
   test(
     `Parler TTS (ggml): ${variant} ${quant} synthesizes 44.1 kHz audio on the target backend`,
-    { timeout: 900000, skip: isMobile },
+    { timeout: 900000, skip: isMobile || (gpuOnly && SKIP_PARLER_GPU_ONLY) },
     async (t) => {
       if (variant === 'large' && os.totalmem() < LARGE_MIN_RAM_BYTES) {
         t.pass(
@@ -542,7 +567,12 @@ for (const { variant, quant } of PARLER_QUANT_MATRIX) {
       }
 
       const useGPU = isApple && !NO_GPU
-      const model = await loadParlerTTS({ parlerModelPath: download.path, seed: 42, useGPU })
+      const model = await loadParlerTTS({
+        threads: TTS_TEST_THREADS,
+        parlerModelPath: download.path,
+        seed: 42,
+        useGPU
+      })
       try {
         const text =
           variant === 'indic'