feat(tts): add Chinese to Kokoro

.cspell-wordlist.txt

@@ -424,3 +424,10 @@ denoiser
 imageembeddings
 unparseable
 zipdepth
+xiaobei
+xiaoxiao
+xiaoyi
+yunjian
+yunxi
+yunxia
+yunyang

apps/benchmarks/scripts/variant-sizes.json

@@ -418,5 +418,14 @@
   "https://huggingface.co/software-mansion/react-native-executorch-clip-vit-base-patch32/resolve/v0.11.0/coreml/clip_vit_base_patch32_text_coreml_int8.pte": 63921048,
   "https://huggingface.co/software-mansion/react-native-executorch-blazeface/resolve/v0.11.0/vulkan/blazeface_vulkan_fp16.pte": 372866,
   "https://huggingface.co/software-mansion/react-native-executorch-all-MiniLM-L6-v2/resolve/v0.11.0/tokenizer.json": 711544,
-  "https://huggingface.co/software-mansion/react-native-executorch-all-MiniLM-L6-v2/resolve/v0.11.0/coreml/all_minilm_l6_v2_coreml_int8.pte": 22817740
+  "https://huggingface.co/software-mansion/react-native-executorch-all-MiniLM-L6-v2/resolve/v0.11.0/coreml/all_minilm_l6_v2_coreml_int8.pte": 22817740,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/phonemizer/zh/lexicon.json": 2334963,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/phonemizer/zh/segmenter.json": 8288194,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zf_xiaobei.bin": 522240,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zf_xiaoxiao.bin": 522240,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zf_xiaoyi.bin": 522240,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zm_yunjian.bin": 522240,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zm_yunxi.bin": 522240,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zm_yunxia.bin": 522240,
+  "https://huggingface.co/software-mansion/react-native-executorch-kokoro/resolve/v0.11.0/voices/zm_yunyang.bin": 522240
 }

apps/speech/app/kokoro-text-to-speech/index.tsx

@@ -25,6 +25,7 @@ const LANGUAGE_OPTIONS = [
   { label: 'Hindi', value: 'HI' as const },
   { label: 'Polish', value: 'PL' as const },
   { label: 'German', value: 'DE' as const },
+  { label: 'Chinese', value: 'ZH' as const },
 ];
 
 type KokoroLanguage = (typeof LANGUAGE_OPTIONS)[number]['value'];
@@ -56,6 +57,7 @@ const SAMPLE_TEXTS: Record<KokoroLanguage, string> = {
   HI: 'कोकोरो एक छोटा टेक्स्ट-टू-स्पीच मॉडल है जो पूरी तरह से आपके डिवाइस पर चलता है।',
   PL: 'Kokoro to niewielki model syntezy mowy, który działa w całości na twoim urządzeniu, bez połączenia z internetem.',
   DE: 'Kokoro ist ein kompaktes Sprachsynthesemodell, das vollständig auf deinem Gerät läuft, ganz ohne Internetverbindung.',
+  ZH: 'Kokoro 是一个小巧的语音合成模型，完全在你的设备上运行，无需联网。',
 };
 // cspell:enable
 

docs/docs/02-extensions/speech/02-text-to-speech.md

@@ -175,7 +175,7 @@ The library provides ready-to-use Text-to-Speech models from the [Software Mansi
 | Model                          | Variants                                                                 | Sub-Models & Assets                                                          | Sample Rate | Size Range  | Supported Backends                           | Supported Languages                                                                                                                                                       | Notes                                                                                                                                                                                                        |
 | :----------------------------- | :----------------------------------------------------------------------- | :--------------------------------------------------------------------------- | :---------- | :---------- | :------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | **Supertonic 3**               | [See](../../06-api-reference/variables/models.md#texttospeechsupertonic) | Text Encoder, Duration Predictor, Vector Estimator, Vocoder, Voice Styles    | 44.1 kHz    | 398 MB      | XNNPACK (CPU), MLX (Apple), Vulkan (Android) | [English, Spanish, French, German, Korean, Japanese & more](../../06-api-reference/react-native-executorch/namespaces/speech/variables/SUPERTONIC_SUPPORTED_LANGUAGES.md) | Faster generation and broad multilingual coverage across 10 bundled [speaker styles](../../06-api-reference/variables/SUPERTONIC_DEFAULT_VOICE_NAMES.md), with slightly lower voice naturalness than Kokoro. |
-| **Kokoro (Language Packages)** | [See](../../06-api-reference/variables/models.md#texttospeechkokoro)     | Duration Predictor, Synthesizer, G2P Lexicon / Neural Model, Language Voices | 24.0 kHz    | 332 MB each | XNNPACK (CPU), Core ML (Apple)               | English (US/GB), Spanish, French, Italian, Portuguese, Hindi, Polish, German                                                                                              | Exceptional voice naturalness and intonation per language package, but heavier compute per synthesized chunk.                                                                                                |
+| **Kokoro (Language Packages)** | [See](../../06-api-reference/variables/models.md#texttospeechkokoro)     | Duration Predictor, Synthesizer, G2P Lexicon / Neural Model, Language Voices | 24.0 kHz    | 332 MB each | XNNPACK (CPU), Core ML (Apple)               | English (US/GB), Spanish, French, Italian, Portuguese, Hindi, Polish, German, Chinese                                                                                     | Exceptional voice naturalness and intonation per language package, but heavier compute per synthesized chunk.                                                                                                |
 
 :::note Model-Specific Pipelines
 Because Text-to-Speech architectures require distinct multi-model orchestration in TypeScript (coordinating phonemizers, duration predictors, flow-matching loops, and neural vocoders), TTS pipelines are model-specific. To use custom voices or fine-tuned weights, provide your modified `.pte` models or custom voice JSON/BIN files matching the [`SupertonicTtsModel`](../../06-api-reference/type-aliases/SupertonicTtsModel.md) or [`KokoroTtsModel`](../../06-api-reference/type-aliases/KokoroTtsModel.md) specifications.

packages/react-native-executorch/cpp/tests/extensions/PhonemizerTest.cpp

@@ -25,7 +25,7 @@ TEST_F(PhonemizerTest, AcceptsEverySupportedLanguage) {
     // The language string is passed straight through to phonemis, which picks
     // the pipeline from it. A profile it does not know throws (below), so this
     // is what catches a name drifting apart from the TS side.
-    for (const auto *lang : {"en-us", "en-gb", "de", "fr", "es", "it", "pl", "pt", "hi"}) {
+    for (const auto *lang : {"en-us", "en-gb", "de", "fr", "es", "it", "pl", "pt", "hi", "zh"}) {
         EXPECT_TRUE(evalBool(std::string("__rnexecutorch_jsi__.speech.createPhonemizer({ lang: '") +
                              lang + "' }); return true;"))
             << "rejected language: " << lang;

packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts

@@ -11,7 +11,17 @@ declare const phonemizerBrand: unique symbol;
  * Union of all supported language codes in the G2P pipeline.
  * @category Speech / Types
  */
-export type PhonemizerLanguage = 'en-us' | 'en-gb' | 'fr' | 'es' | 'it' | 'pt' | 'de' | 'pl' | 'hi';
+export type PhonemizerLanguage =
+  | 'en-us'
+  | 'en-gb'
+  | 'fr'
+  | 'es'
+  | 'it'
+  | 'pt'
+  | 'de'
+  | 'pl'
+  | 'hi'
+  | 'zh';
 
 /**
  * Configuration options and asset paths for initializing a {@link Phonemizer}.
@@ -51,9 +61,10 @@ export type PhonemizedText = {
   /** IPA transcription of the whole text, the same as `phonemize(text)` returns. */
   readonly phonemes: string;
   /**
-   * The whitespace-separated words of the input text, in order. Each maps onto
-   * one space-separated phoneme group, except numbers, which map onto as many as
-   * they were spelled out into. Empty if the words could not be matched.
+   * The whitespace-separated words of the input text (segmented words for
+   * Chinese), in order. Each maps onto one space-separated phoneme group,
+   * except numbers, which map onto as many as they were spelled out into.
+   * Empty if the words could not be matched.
    */
   readonly words: readonly PhonemizedWord[];
 };

packages/react-native-executorch/src/models.ts

@@ -1848,9 +1848,9 @@
   synthesizer: `${KOKORO_ROOT}/${backend}/${dir}/synthesizer_${variant}_${backend}_fp32.pte`,
 });
 
-const kokoroVoices = <const N extends string>(names: readonly N[]) =>
+const kokoroVoices = <const N extends string>(names: readonly N[], root = KOKORO_ROOT) =>
   names.reduce(
-    (acc, name) => ({ ...acc, [name]: `${KOKORO_ROOT}/voices/${name}.bin` }),
+    (acc, name) => ({ ...acc, [name]: `${root}/voices/${name}.bin` }),
     {} as Record<N, string>
   );
 
@@ -1863,13 +1863,31 @@
   neuralModelSource: `${KOKORO_PHONEMIZER_ROOT}/${lang}/phonemizer_${lang.replace('-', '_')}.pte`,
 });
 
-const kokoroNeuralPhonemizer = <const L extends Exclude<PhonemizerLanguage, 'en-us' | 'en-gb'>>(
+const kokoroNeuralPhonemizer = <
+  const L extends Exclude<PhonemizerLanguage, 'en-us' | 'en-gb' | 'zh'>,
+>(
   lang: L
 ) => ({
   lang,
   neuralModelSource: `${KOKORO_PHONEMIZER_ROOT}/${lang}/phonemizer_${lang}.pte`,
 });
 
+// Chinese is published from the next tag. Its phonemizer segments the text
+// into words, with the segmenter data in the tagger slot, and reads each word
+// from the lexicon. Upstream's zf_xiaoni is left out: it is unintelligible even
+// in the reference PyTorch model.
+const KOKORO_NEXT_ROOT = `${BASE_URL}-kokoro/${NEXT_VERSION_TAG}`;
+const KOKORO_ZH_PHONEMIZER = {
+  lang: 'zh' as const,
+  taggerSource: `${KOKORO_NEXT_ROOT}/phonemizer/zh/segmenter.json`,
+  lexiconSource: `${KOKORO_NEXT_ROOT}/phonemizer/zh/lexicon.json`,
+};
+const KOKORO_ZH_VOICES = kokoroVoices(
+  ['zf_xiaobei', 'zf_xiaoxiao', 'zf_xiaoyi', 'zm_yunjian', 'zm_yunxi', 'zm_yunxia', 'zm_yunyang'],
+  KOKORO_NEXT_ROOT
+);
+type KokoroZhVoice = keyof typeof KOKORO_ZH_VOICES;
+
 const KOKORO_EN_US_XNNPACK_FP32: KokoroTtsModel<
   'af_heart' | 'af_river' | 'af_sarah' | 'am_adam' | 'am_michael' | 'am_santa'
 > = {
@@ -1926,6 +1944,12 @@
   phonemizer: kokoroNeuralPhonemizer('de'),
   voices: kokoroVoices(['df_anna']),
 };
+const KOKORO_ZH_XNNPACK_FP32: KokoroTtsModel<KokoroZhVoice> = {
+  name: 'kokoro',
+  modelPaths: kokoroModelPaths('xnnpack', 'std', 'standard'),
+  phonemizer: KOKORO_ZH_PHONEMIZER,
+  voices: KOKORO_ZH_VOICES,
+};
 
 // Core ML counterparts: the same weights, phonemizer and voices as the XNNPACK
 // presets, only the `.pte` files differ. Their token axis is fixed at 128, so
@@ -1987,6 +2011,12 @@
   phonemizer: kokoroNeuralPhonemizer('de'),
   voices: kokoroVoices(['df_anna']),
 };
+const KOKORO_ZH_COREML_FP32: KokoroTtsModel<KokoroZhVoice> = {
+  name: 'kokoro',
+  modelPaths: kokoroModelPaths('coreml', 'std', 'standard'),
+  phonemizer: KOKORO_ZH_PHONEMIZER,
+  voices: KOKORO_ZH_VOICES,
+};
 
 // =============================================================================
 // Privacy Filter
@@ -3622,6 +3652,10 @@
         XNNPACK_FP32: KOKORO_DE_XNNPACK_FP32,
         COREML_FP32: KOKORO_DE_COREML_FP32,
       }),
+      ZH: variants({
+        XNNPACK_FP32: KOKORO_ZH_XNNPACK_FP32,
+        COREML_FP32: KOKORO_ZH_COREML_FP32,
+      }),
     },
   },