feat: add multilingual TTS support via ephone (eSpeak-NG WASM)

[Dongshan-git]

Summary

• Replace the English-only phonemizer package with ephone v1.0.2, an eSpeak-NG WASM wrapper with built-in language packs for 9 languages (en-US, en-GB, es, fr, it, pt-BR, ja, zh, hi)
• Fix a hang in KokoroTTS.stream() where splitter.close() was never called, leaving the async iterator blocked forever after all chunks were pushed
• Uncomment all non-English voices in voices.js (Japanese, Chinese, Spanish, French, Hindi, Italian, Portuguese)
• Redesign the browser demo with a grouped voice selector, waveform visualizer, and per-language example texts

Changes

src/phonemize.js

• Rewrite around createEphone from the ephone package
• Language packs (en_us, en_all, roa, jpx, sit) are imported as static top-level imports; Hindi's large all pack is lazy-loaded on demand
• normalize_text now takes an english flag — number/currency/abbreviation normalization is skipped for non-English input
• English post-processing (r → ɹ, kokoro pronunciation fix) is gated behind isEnglish so it doesn't corrupt romance-language phonemes

src/kokoro.js

• Add missing splitter.close() call after splitter.push(...chunks) — without this the TextSplitterStream async iterator never resolves
• Await this.tokenizer(...) (was missing await, causing silent failures)
• Extend _validate_voice language type annotation to cover all 9 language codes

src/voices.js

• Enable all previously-commented-out non-English voices

rollup.config.js

• Switch web build from file: "kokoro.web.js" to dir + entryFileNames/chunkFileNames to support dynamic imports (ephone language packs are code-split chunks)

demo/

• Grouped voice selector with language flags and quality grades
• WaveformPlayer component with canvas waveform visualization
• Per-language example texts
• AnimatePresence loading/result transitions

Test plan

• English (af_heart, bf_emma) — verify existing behaviour unchanged
• French (ff_siwis) — was completely broken before; should now synthesize correctly
• Japanese (jf_alpha) — hiragana/katakana input; kanji not supported by eSpeak-NG
• Chinese (zf_xiaoxiao), Spanish (ef_dora), Hindi (hf_alpha), Italian (if_sara), Portuguese (pf_dora)
• Run vitest — all phonemize tests should pass

🤖 Generated with Claude Code

[Dongshan-git]

Also bumps @huggingface/transformers from ^3.5.1 to ^4.0.1 and dev dependencies (rollup, vitest, typescript, prettier) to their latest major versions.