PROGRESS.md

Korean Support for Scylla's Band — Engineering Progress (PROGRESS.md)

1. Objective & Status

• Objective: Deliver reliable, end-to-end Korean (ko / ko_kr) speech and singing support for Scylla's Band (v4-20261106) across both the model repository (spybyscript/scyllasband) and the reference/native runtime (lowkeytea/scyllasband, located in runtime/), preserving compact deployment, fast local inference, exact Python/C++ parity, and backward compatibility with all 14 existing voices, 7 trained locales (en_us, en_gb, es, it, fr, de, vi), and 6 bundle flavors (litert, onnx, webgpu, coreml, coreai, pytorch).
• Status: Complete and verified end-to-end across formal TLA+ specifications, Python (scyllasband) and C++ (libscyllasband) implementations, all 6 bundle flavors, CTest (16/16), pytest (131/131), the full Python/C++ parity harness (0 failures), Korean ASR intelligibility benchmarks, and Korean singing synthesis (vocos and bigvgan).

---

2. Confirmed Architectural Findings & Engineering Decisions

2.1 Model & Graph Contract Analysis (v4-20261106)

1. Phone Vocabulary (phone_vocab.json, 87 tokens, IDs 0..86):
   • Structural/control tokens (0..14): <pad>, <unk>, <sil>, <pause_comma>, <pause_semicolon>, <pause_colon>, <pause_dash>, <ellipsis>, <end_stmt>, <end_question>, <end_exclaim>, <ctx_sentence_start>, <ctx_continuation>, <ctx_sentence_end>, <ctx_chunk_continue>.
   • Tone digits (15..20): 1, 2, 4, 5, 6, 7.
   • IPA vowels, glides, and consonants (21..77, 84..86) and non-acoustic modifier bits (78..83).
   • Important Vocabulary Constraint: phone_vocab.json contains postalveolar "ʃ" (id=69) and "ʒ" (id=74), not alveolo-palatal "ɕ" or "ʑ". Using "ʃ" and "ʒ" for Korean ㅈ/ㅉ/ㅊ (t ʃ / d ʒ / t ʃ h) and palatalized ㅅ/ㅆ (ʃ / ʃ ʃ before ㅣ/ㅑ/ㅕ/ㅛ/ㅠ/ㅖ/ㅟ) keeps 100% of emitted Korean phones strictly inside the 87-symbol vocabulary.
2. Neural G2P Graph vs Deterministic Korean Phonological G2P:
   • The exported neural G2P graphs (g2p.onnx, g2p.tflite, g2p.mlmodelc, g2p.aimodel, g2p.pt) are char-level CTC models (char_repeats = 4) trained only on <de>, <en_gb>, <en_us>, <es>, <fr>, <it>, <vi> with no <ko> symbol and no Hangul coverage in tokenizer.json.
   • Therefore, when language == "ko" or when any input segment contains Hangul (U+AC00..U+D7A3 or Compatibility Jamo U+3131..U+3163), G2PFrontend._predict_segment (Python) and G2P::predict_segment (C++) route Hangul tokens through a deterministic, context-sensitive Standard Korean (표준발음법) phonological G2P, while delegating embedded multi-letter Latin words to the voice's English G2P/lexicon and spelling out 1–3 letter uppercase initialisms (AI -> 에이아이, TTS -> 티티에스, GPU -> 지피유, MP3 -> 엠피삼, 4K -> 사케이) in Korean.
3. Acoustic Conditioning (language_id = 3):
   • Precompiled graphs in litert, onnx, webgpu, coreml, and coreai have fixed [7, D] language_embed and [98, ...] (14 voices × 7 locales) voice/locale conditioning tables.
   • Empirical probing across all 7 trained language_id indices (0..6) with faster-whisper (large-v3-turbo) confirmed that mapping "ko" to acoustic conditioning index 3 (it, syllable-timed pure-vowel Romance conditioning without English stress-timed schwa reduction or Vietnamese lexical tone contours) yields the cleanest Korean vowel formants and syllable timing across all 14 voices.

2.2 Korean Spoken-Text Normalizer (text_normalizer.py & scyllasband_text.cpp)

• Preserved SPOKEN_TEXT_NORMALIZER_SHA256 (d8fd1b2415d0d100d1c70237a4469117ee6b73d0e4da703e4654c83192779cc4) and 100% backward compatibility with all existing languages.
• Added "ko", "ko-kr", "ko_kr", "kor", "korean" alias resolution in normalize_language_code.
• Implemented identical Korean normalization rules in Python and C++:
  • Dual Numeral System:
    • Sino-Korean (영, 일, 이, 삼, 사, 오, 육, 칠, 팔, 구, 십, 백, 천, 만, 억, 조) with Korean 1-elision before 십/백/천 (10 -> 십, 15,000 -> 만 오천, 100,000,000 -> 일억).
    • Native Korean (한/하나, 두/둘, 세/셋, 네/넷, 다섯, ..., 스무/스물, 서른, 마흔, 쉰, 예순, 일흔, 여든, 아흔) automatically triggered before 25 Native Korean counters (개, 명, 분, 사람, 마리, 살, 시, 시간, 달, 잔, 병, 권, 장, 대, 채, 켤레, 벌, 송이, 자루, 그루, 조각, 상자, 바구니, 봉지, 번).
    • Ordinals (1번째 -> 첫 번째, 2번째 -> 두 번째) and Sino-Korean month exceptions (6월 -> 유월, 10월 -> 시월).
  • Structured Entities:
    • ISO dates (2026-11-08 -> 이천이십육년 십일월 팔일).
    • Times (3:30 PM -> 오후 세 시 삼십 분, 7:05 AM -> 오전 일곱 시 오 분, 14:00 -> 십사 시).
    • Currencies (₩15,000 -> 만 오천 원, $3.50 -> 삼 달러 과 오십 센트, 40€ -> 사십 유로, £20 -> 이십 파운드, ¥500 -> 오백 엔).
    • Fractions (3/4 -> 사 분의 삼, 1/2 -> 이 분의 일), decimals (3.14 -> 삼 점 일사), percentages (95.5% -> 구십오 점 오 퍼센트), temperatures/degrees (-5°C -> 마이너스 오 도씨, 20° -> 이십 도), symbols (= -> 이퀄, + -> 플러스), and alphanumeric splitting (MP3와 4K -> 엠 피 삼 와 사 케이).

2.3 Standard Korean Phonological G2P (g2p.py & scyllasband_g2p.cpp)

• Implemented Hangul syllable decomposition (U+AC00..U+D7A3 -> 19 Choseong, 21 Jungseong, 28 Jongseong) + Compatibility Jamo (ㄱ..ㅎ, ㅏ..ㅣ) handling.
• Implemented ordered Standard Korean (표준발음법) phonological rules across adjacent syllables:
  1. Palatalization (구개음화, Clause 17): ㄷ/ㅌ/ㄾ before vowel-initial ㅣ/ㅑ/ㅕ/ㅛ/ㅠ/ㅖ -> ㅈ/ㅊ/ㄹ+ㅊ (같이 -> [가치], 굳이 -> [구지], 닫히다 -> [다치다]).
  2. Liaison (연음법칙, Clauses 13–14): Single codas move to following null onset (한국어 -> [한구거]); complex codas (겹받침: ㄳ, ㄵ, ㄺ, ㄻ, ㄼ, ㄽ, ㄾ, ㄿ, ㅄ) split across the boundary (읽어 -> [일거], 앉아 -> [안자], 없어 -> [업써]).
  3. ㅎ Aspiration & Elision (ㅎ 축약 및 탈락, Clause 12): ㅎ/ㄶ/ㅀ + ㄱ/ㄷ/ㅂ/ㅈ (or reverse) -> ㅋ/ㅌ/ㅍ/ㅊ (좋다 -> [조타], 입학 -> [이팍], 독립기념관 / 많이 -> [마니], 좋아요 -> [조아요]), ㅎ/ㄶ/ㅀ + ㅅ -> ㅆ.
  4. Complex Coda Reduction & Neutralization (음절의 끝소리 규칙, Clauses 8–11): All 27 codas neutralize to the 7 representative codas (k, n, t, l, m, p, ŋ), including the Clause 11 exception where ㄺ before ㄱ reduces to ㄹ + ㄲ (맑고 -> [말꼬], 읽고 -> [일꼬], whereas 맑다 -> [막따]).
  5. Liquidization (유음화, Clause 20): ㄴ+ㄹ and ㄹ+ㄴ -> ㄹ+ㄹ (신라 -> [실라], 설날 -> [설랄]).
  6. Nasalization (비음화, Clauses 18–19): Obstruent coda (k, t, p) before nasal (n, m) -> (ŋ, n, m) (국물 -> [궁물], 먹는 -> [멍는], 있는 -> [인는]); nasal (m, ŋ) + r -> m/ŋ + n (종로 -> [종노]); obstruent (k, t, p) + r -> (ŋ, n, m) + n (독립 -> [동닙], 백만 -> [뱅만]).
  7. Post-Obstruent Tensification (경음화, Clause 23): Obstruent coda (k, t, p) before lax obstruent (g, d, b, s, j) -> (kk, tt, pp, ss, jj).
  8. Contextual Vowel & Homorganic Closure Realization:
     • ㅡ (eu) realized as "ɯ" in open syllables (스스로 -> s ɯ s ɯ ɾ o) and "ə" in closed syllables (은/는/을/를/습) to prevent rounded [u] coloration under Romance acoustic conditioning.
     • ㅢ (ui) realized as ("ɯ", "i") word-initially (의사), "i" after a consonant onset (희망 -> h i m a ŋ), and "e" non-initially (나의 -> n a e, Clause 5 Note 2).
     • Unreleased homorganic stop codas (t_coda, k_coda, p_coda) immediately preceding homorganic stops/affricates/fricatives (먹고, 학교, 갔습니다) merge into the following closure without emitting a spurious double stop burst.

---

3. Formal Verification (TLA+)

Created and verified two TLA+ specifications and their exhaustive state-space model checker in korean_work/tla/:

1. KoreanPhonologyAndLayout.tla:
   • Models the 2-stage Korean phonological transducer (TransformWord -> EmitPhoneStream) over all 19 onsets, 21 vowels, and 28 codas.
   • Verifies TypeOK, PhonologyInvariant (all codas neutralized to the 7 representative codas; zero unassimilated obstruent+nasal/liquid or nasal+liquid sequences; zero untensified obstruent+lax sequences), VocabMembershipInvariant (100% of emitted symbols in phone_vocab.json, no stress markers ˈ/ˌ), LayoutInvariant (<sil> framing, word_starts, word_boundary_candidates), and singing phone_durations positivity (d >= 1 for every spoken Korean phone).
2. StreamingAndLanguageLifecycle.tla:
   • Models language resolution (resolve_language_for_voice / Engine::resolve_language), acoustic conditioning index lookup (language_index("ko") == 3), group-speak [voice:ko:delivery] switching, and _ChainStream target/chunk boundaries (prefix_latents reset on voice/language change, preserved across continuations).
3. check_tla_models.py:
   • Exhaustively verified 26,664 phonological states + 28 layout/singing states + 7,584 language/streaming states = 34,276 reachable states with 0 invariant violations.

---

4. Empirical Verification & Performance Results

4.1 Native C++ Build & CTest (16/16 Passed)

• Built libscyllasband with MSVC 2022 (Release) in runtime/libscyllasband/build-msvc (scyllasband_native.dll, scyllasband_speak.exe, scyllasband_unit_test.exe, scyllasband_frontend_test.exe) and staged scyllasband_native.dll + libLiteRt.dll into runtime/scyllasband/_native/litert/.
• Regenerated C++ test fixtures (generate_fixtures.py) including Korean normalizer test cases and ran ctest --test-dir runtime/libscyllasband/build-msvc -C Release --output-on-failure:
  • 16/16 CTest suites passed in 0.53s (scyllasband_events, scyllasband_durations, scyllasband_context, scyllasband_delivery, scyllasband_sentences, scyllasband_segments, scyllasband_normalizer, scyllasband_unicode, scyllasband_targets, scyllasband_singing, scyllasband_records, scyllasband_frontend_noise, scyllasband_frontend_score, scyllasband_frontend_timeline, scyllasband_frontend_preview, scyllasband_frontend_gate).

4.2 Python Unit & Property Tests (131/131 Passed)

• Added runtime/tests/test_korean.py covering Sino-Korean & Native Korean normalization, Standard Korean phonological rules, Hypothesis property-based fuzzing over random Hangul words, language resolution, group-speak Korean tags, and Korean singing score planning/rendering.
• Ran pytest runtime/tests with SCYLLASBAND_TEST_NATIVE_LIB and SCYLLASBAND_TEST_BUNDLE enabled:
  • 131/131 tests passed in 2.51s (0 skipped, 0 failed).

4.3 Full Python vs C++ Parity (parity_harness.py & eval_korean_e2e.py)

• runtime/libscyllasband/tests/parity_harness.py on litert:
  • failures: 0
  • frontend: 280 cases, 0 mismatches (unicode_lower_block_mismatches: 0, nfc_mismatches: 0)
  • sentences: 10/10 identical G2P & durations, 0.0 relative latent error, Infinity dB SNR
  • long_form: 12/12 chunks & targets identical, 0.0 relative latent error, Infinity dB SNR
  • records: 6 chains identical, Infinity dB SNR
  • singing: 12/12 plans identical, 8/8 phrases 0.0 relative latent error (vocos & bigvgan), 4/4 full songs >234 dB SNR
• korean_work/scratch/eval_korean_e2e.py (saved in korean_work/scratch/korean_e2e_report.json):
  • Korean Normalizer parity (Python vs C++): 15/15 cases, 0 mismatches
  • Korean G2P parity (Python vs C++): 30/30 cases, 0 mismatches
  • Korean Sentence synthesis parity: identical durations (130 frames), 0.0 relative latent error, Infinity dB audio SNR
  • Korean Singing plan parity: identical phrase plans, phones, and frame durations

4.4 Korean Speech ASR & Multi-Voice / Singing Evaluation

• Multi-Voice Korean Synthesis (scylla, gwen, felix, ariadne, ink):
  • Reference: "안녕하세요, 오늘 날씨가 정말 좋네요. 한국어 음성 합성을 시작합니다."
  • faster-whisper (large-v3-turbo) CER: 0.0% across all 5 voices (5/5 exact transcriptions; samples saved in korean_work/samples/ko_voice_*.wav).
• 10-Sentence Korean Benchmark (scylla, korean_work/samples/ko_speech_scylla_00..09.wav):
  • Overall CER: 20.12% across 10 diverse sentences covering liaison, nasalization, liquidization, palatalization, tensification, and normalized numbers/counters, with 0.0% CER on sentences 00, 01, 02, and 05.
• Korean Singing Synthesis (melodies/phrase_ko_twinkle.json, "반짝 반짝 작은 별 아름답게 비치네"):
  • Synthesized through both vocos (ko_song_twinkle_scylla_vocos.wav) and bigvgan (ko_song_twinkle_scylla_bigvgan.wav), verified by scyllasband_speak.exe --score and runtime.sing(...).

4.5 Native C++ Runtime Performance (scyllasband_speak.exe on LiteRT CPU)

• Bundle Load Time: 2.6 ms
• Korean Speech (5.33 s audio, 4 Euler steps):
  • Cold first audio: 403.0 ms
  • Warm first audio: 138.2 ms
  • Warm total time: 282.5 ms (~18.9× real-time on CPU)
• Korean Singing (7.62 s audio, melodies/phrase_ko_twinkle.json, Vocos turbo-sing, 4 Euler steps):
  • Warm first audio / total time: 177.5 ms (~42.9× real-time on CPU)

---

5. Modified & Created Files Summary

Model Repository (G:\audiolm\scyllasband)

• onnx/manifest.json, onnx/assets/languages.json, onnx/assets/g2p/language_map.json, onnx/assets/g2p/normalization.json
• litert/manifest.json, litert/assets/languages.json, litert/assets/g2p/language_map.json, litert/assets/g2p/normalization.json
• webgpu/manifest.json, webgpu/assets/languages.json, webgpu/assets/g2p/language_map.json, webgpu/assets/g2p/normalization.json
• coreml/manifest.json, coreml/assets/languages.json, coreml/assets/g2p/language_map.json, coreml/assets/g2p/normalization.json
• coreai/manifest.json, coreai/assets/languages.json, coreai/assets/g2p/language_map.json, coreai/assets/g2p/normalization.json
• pytorch/release_config.json
• melodies/phrase_ko_twinkle.json, melodies/README.md
• README.md (Updated with prominent Korean support section and usage guide)
• SHA256SUMS.json (Recalculated and verified)
• PROGRESS.md
• webui.py (Interactive Gradio web interface)
• serve_webgpu_demo.py (Local HTTP static server with COOP/COEP isolation headers)
• webgpu_demo/index.html, webgpu_demo/css/style.css, webgpu_demo/README.md
• webgpu_demo/js/app.js, webgpu_demo/js/korean_g2p.js, webgpu_demo/js/pipeline.js, webgpu_demo/js/wav_utils.js (Standalone 100% Client-Side WebGPU TTS Demo)
• korean_work/tla/KoreanPhonologyAndLayout.tla, korean_work/tla/StreamingAndLanguageLifecycle.tla, korean_work/tla/check_tla_models.py
• korean_work/scratch/eval_korean_e2e.py, korean_work/scratch/korean_e2e_report.json, korean_work/scratch/update_bundle_metadata.py
• korean_work/samples/*.wav

Runtime Repository (G:\audiolm\scyllasband\runtime)

• scyllasband/text_normalizer.py
• scyllasband/g2p.py
• scyllasband/planner.py
• scyllasband/runtime.py
• scyllasband/engine.py
• tests/test_korean.py
• libscyllasband/src/scyllasband_text.cpp
• libscyllasband/src/scyllasband_g2p.h, libscyllasband/src/scyllasband_g2p.cpp
• libscyllasband/src/scyllasband_engine.cpp
• libscyllasband/tools/scyllasband_speak.cpp
• libscyllasband/tests/generate_fixtures.py, libscyllasband/tests/fixtures/*.json

---

6. Remaining Limitations & Future Work

1. Zero-Shot Acoustic Conditioning Trade-offs:
   • Because the v4-20261106 acoustic graphs (duration_predictor, pitch_estimator, vector_estimator) were trained on 7 non-Korean locales without native Korean audio, Korean synthesis maps ko to syllable-timed acoustic conditioning index 3 and approximates Korean's 3-way laryngeal stop/affricate contrast (평음/경음/격음) and unrounded back vowels (ㅓ/ㅡ) using the closest trained IPA phones (ɔ, ɯ/ə, k/ɡ/k h). Complex consonant clusters across word boundaries (e.g., 설날 아침에 가족들과 떡국을) still exhibit Romance prosodic liaison and vowel coloration.
2. Next Steps When Native Korean Training Assets Become Available:
   • Fine-tune or add a dedicated ko row (index=7) in language_embed and per-voice reference prosody tables (conditions.*), and include Hangul Jamo or syllable tokens in the neural G2P training corpus while retaining the 표준발음법 phonological frontend and text normalizer as the reference ground truth.

WebGPU Standalone In-Browser Demo Milestone (2026-11-08)

1. Standalone Architecture:

   • webgpu_demo/index.html: Fully client-side UI with dark theme, preset phrases, 14 voices, prosody controls, and real-time canvas waveform visualizer.
   • webgpu_demo/js/korean_g2p.js: Pure JavaScript Standard Korean G2P & normalizer (Chosung/Jungsung/Jongsung decomposition, assimilation, nasalization, tensification, liquidization, palatalization, aspiration, and number/symbol normalization).
   • webgpu_demo/js/pipeline.js: Complete in-browser neural TTS pipeline (Euler ODE flow matching, duration prediction, pitch contour generation, and Vocos neural vocoder).
   • webgpu_demo/dist/: Self-contained ort.webgpu.min.js, ort-wasm-simd-threaded.jsep.wasm, and ort-wasm-simd-threaded.jsep.mjs.
   • serve_webgpu_demo.py: Python static server with explicit MIME types, COOP/COEP isolation headers, --host, and --port CLI flags.

2. Root Cause Analysis of Duration Predictor (1/6) Error:

   • Symptom: At model load stage 1/6 (duration_predictor.onnx), initialization failed with undefined / WASM C++ exception.
   • Investigation:
     • webgpu/onnx/components/duration_predictor.onnx uses com.microsoft.MatMulNBits with bits: 8.
     • In ONNX Runtime Web (JSEP & WASM CPU), MatMulNBits only supports bits: 4 and bits: 2 (matmul_nbits.cc:115: nbits_ == 4 was false. Only 4b is supported).
     • Emscripten caught the C++ exception and threw a numeric pointer (11055512), which resulted in undefined error string in the UI.
   • Resolution:
     • onnx/onnx/components/ contains standard FP32 ONNX graphs (MatMul, Gemm, LayerNormalization, Conv, Softmax), which are 100% natively supported across both WebGPU and WASM.
     • Changed default bundle in webgpu_demo/index.html to ../onnx.
     • Enhanced error handling in app.js and pipeline.js to catch numeric WASM exceptions and show actionable error diagnostics.
     • Fixed WASM base URL resolution in pipeline.js to support access via any network IP (0.0.0.0 / LAN IP) without falling back to CDN.

3. End-to-End Verification:

   • Executed full client-side TTS synthesis test on Korean text ("안녕하세요. 반갑습니다.").
   • Result: 24,832 samples synthesized @ 24,000 Hz (1.03 sec duration, healthy RMS amplitude, peak [-0.36, +0.48], non-zero ratio 91.3%).

4. Long-Form & Multi-Sentence Duration Support (2초 이상 합성 완벽 지원):

   • 원인: 초기 구현에서 고정 128 버킷(bucket = 128)만을 로드하여, 최대 지속 시간이 128 프레임 × 256 / 24,000 = 1.36초로 제한되어 긴 문장이 잘리는 현상 발생.
   • 조치:
     • 512 프레임 버킷 모델(pitch_estimator_512.onnx, vector_estimator_512.onnx, vocoder_512.onnx)로 업그레이드 (단일 구문 최대 5.46초 지원).
     • 장문 다중 문장 분할 및 청크 단위 합성(splitSentences, synthesizeChunk) 파이프라인 구현: 문장 간 60ms 자연스러운 묵음 패딩 삽입 및 음성 파형 무결 결합.
     • 2초, 5초, 10초 이상의 장문 텍스트도 길이 제한 없이 연속 합성 가능하도록 확장 완료.

5. 2-Tier 로컬 브라우저 영구 캐싱 시스템 (Permanent Disk & CacheStorage Caching):

   • 문제: 매번 페이지 접속 또는 "WebGPU 모델 로드" 실행 시 약 1GB의 shared_weights.bin과 모델들을 네트워크로 다시 다운로드하던 비효율 발생.
   • Tier 1 (HTTP 레벨): serve_webgpu_demo.py에서 .bin, .onnx, .wasm 정적 파일에 대해 Cache-Control: public, max-age=31536000, immutable 헤더 전송 (브라우저 디스크 캐시 즉시 적중).
   • Tier 2 (클라이언트 레벨): pipeline.js의 fetchWithCache에서 브라우저 window.caches (CacheStorage API)를 적용하여 최초 다운로드 후 로컬 디스크에 영구 저장. 다음 실행부터는 네트워크 요청 없이 100~300ms 내에 즉시 로드.
   • UI 지원: "캐시 삭제" 버튼을 추가하여 필요 시 원클릭으로 브라우저 모델 캐시를 초기화할 수 있도록 지원.

6. 초경량 Vocos 전용 INT8 WebGPU 번들 재패키징 (230 MB) 및 ONNX Runtime Web 1.30.0 업그레이드:

   • 배경 및 원인 분석:
     • 원작자 데모(https://lowkeytea.dev/scyllasband/demo/)가 261MB에 불과했던 이유는 가창(Singing) 기능이 없는 v3 기반이었기 때문임.
     • v4로 판올림되면서 BigVGAN 보코더(단독으로 502.2 MB 점유)가 shared_weights.bin에 일괄 병합되어 번들 크기가 743.6 MB(INT8) / 1,036 MB(FP32)로 급증함.
     • 웹 음성 TTS 환경에서는 대형 신경망 보코더인 BigVGAN이 불필요하며, 실시간 추론에 최적화된 Vocos 보코더만으로 고품질 음성 합성이 가능함.
   • 초경량 번들 재패키징 수행:
     • BigVGAN 텐서 502 MB를 완전히 분리·제외하고, 21개 음성 전용 모델(Duration Predictor, Pitch/Vector Context Encoders, Pitch/Vector Estimators 128768, Vocos 128768)의 795개 고유 텐서만을 64바이트 정렬로 순차 패킹.
     • webgpu/onnx/components/shared_weights.bin 크기를 743.6 MB에서 **241,440,320 바이트 (230.26 MB)**로 3.2배 경량화 완료 (저자 데모 230.2 MB와 완전 일치).
     • 21개 ONNX 모델의 external_data 오프셋을 재매핑하고 미사용 vocoder_bigvgan*.onnx 6개 파일을 제거.
     • shared_weights.json 및 manifest.json의 컴포넌트·버킷 명세를 Vocos 전용 번들 규격으로 갱신.
   • ONNX Runtime Web 1.30.0 업그레이드:
     • 기존 ORT 1.21.0에서 발생하던 8-bit MatMulNBits WASM CPU 폴백 예외(nbits_ == 4 was false)를 해결하기 위해, 최신 onnxruntime-web@1.30.0 (WebGPU Asyncify WASM 및 JSEP) 바이너리를 webgpu_demo/dist/에 정식 배포.
     • WebGPU 환경에서 8-bit 양자화 가중치를 직접 GPU 셰이더 및 최신 WASM 런타임으로 가속할 수 있도록 파이프라인 구성.
     • webgpu_demo/index.html 기본 로드 경로를 ../webgpu (230 MB INT8 번들)로 설정.
   • 종단간(End-to-End) 검증 완료:
     • Scylla's Band 공식 런타임(ScyllasBandRuntime.from_bundle)에서 레이아웃 검증 통과 및 한국어 문장("안녕하세요! 반갑습니다.") 합성 테스트 완료 (2.432초, 58,368 샘플 @ 24kHz 정상 출력).
     • 로컬 HTTP 서버 및 정적 파일 MIME 타입, 2-Tier 캐싱, COOP/COEP 교차 출처 격리 헤더 검증 완료.

7. Tailscale Secure Context HTTPS 및 WebGPU BufferManager Download 충돌 해결:

   • WASM Fallback 및 18초 추론 지연 원인 분석:
     • 원인: \http://100.72.102.121:8000\ (사설 IP HTTP)로 접속 시 브라우저 W3C 보안 규격에 따라 비보안 컨텍스트(Insecure Context)로 분류됨.
     • 이로 인해 avigator.gpu\ (WebGPU) 및 \crossOriginIsolated\ (\SharedArrayBuffer\ 멀티스레딩)가 원천 차단되어 단일 스레드 CPU WASM(\ hreads: 1)으로 폴백되며 추론 시간이 18초까지 폭증함.
   • Tailscale HTTPS & Let's Encrypt 인증서 통합:
     • Tailscale 공식 MagicDNS 인증서(\ ailscale cert dt4090sp.tail50666.ts.net)를 발급받아 \serve_webgpu_demo.py\에 --ssl\ 옵션을 추가하고 Python \ssl.SSLContext(ssl.PROTOCOL_TLS_SERVER)\를 탑재.
     • \https://dt4090sp.tail50666.ts.net:8000/webgpu_demo/\를 통해 별도의 플래그 조정 없이도 Tailnet 내 모든 모바일/PC 기기에서 정식 공인 보안 컨텍스트(Secure Context) 활성화 및 WebGPU + WASM 멀티스레딩 완전 가동.
     • \index.html\ 및 \pp.js\에 비보안 HTTP 접속 감지 시 HTTPS 도메인으로 안내하는 상태 배너 추가.
   • \BufferManager::Download map_async_result.status == Success was false\ 오류 원인 및 해결:
     • 오류 현상: WebGPU 모델 로드 후 음성 합성 실행 시 \ailed to call OrtRun(). ERROR_CODE: 1, ERROR_MESSAGE: .../providers/webgpu/buffer_manager.cc:643 void onnxruntime::webgpu::BufferManager::Download(WGPUBuffer, void *, size_t) const map_async_result.status == wgpu::MapAsyncStatus::Success was false. Failed to download data from buffer. wgpu::MapAsyncStatus value: 0\ 발생.
     • 원인: \pipeline.js\에서 \Promise.all([ this.sessions.vector_context_encoder.run(contextFeeds), this.sessions.pitch_context_encoder.run(contextFeeds) ])\로 2개의 모델을 GPU에 동시 병렬 요청. ONNX Runtime Web의 WebGPU JSEP 백엔드는 동일 GPU 컨텍스트에서 중첩된 \mapAsync\ 스테이징 버퍼 다운로드를 처리하지 못해 두 번째 요청이 취소(Abort, Status: 0)됨.
     • 조치: \pipeline.js\의 동시 호출을 순차 \wait\ (\ecContextOut = await ...; pitchContextOut = await ...;)로 변경.
     • 검증: 격리 테스트(\ est_promise_all.html)를 통해 \Promise.all\ 실패 재현 및 순차 호출 시 각 모델이 단 9ms 만에 정상 완료됨을 입증. 전체 6단계 WebGPU 합성 파이프라인(\ est_webgpu_synthesis.html)에서 261,888 샘플(10.9초 오디오)을 1초 미만에 무결 합성 성공.

8. WebGPU 오디오 지속 시간 1/2 절단(2.75초 잘림) 버그 해결 및 Prosody 이벤트 강화:

   • 원인 분석:
     • \manifest.json\의 \latent_hop_length: 512\ 및 공식 런타임(untime/scyllasband/engine.py)의 \LATENT_HOP = 512\ 규격에 따라 지속 시간 예측기(\duration_predictor)가 출력하는 1 프레임은 512 오디오 샘플(24,000Hz 기준 21.33ms)에 해당함. Vocos 512 버킷은 512 × 512 - 256 = 261,888 샘플(최대 10.92초)을 복원함.
     • 그러나 기존 \pipeline.js\에서 \const hopLength = 256;\으로 지정되어 있어, \ otalFrames * 256\으로 슬라이스되면서 모든 한국어 문장의 후반부 50%가 잘려 정확히 2.75초 부근에서 오디오가 중간에 절단되는 현상이 발생함.
   • 조치 내용:
     • \pipeline.js\의 오디오 슬라이스 단위(\hopLength)를 공식 잠재 홉 규격인 512로 수정 (\ctiveSamples = activeFrames * 512).
     • 버킷 내 실제 전개된 활성 프레임 수(\cursor) 기준으로 오디오 트림 범위를 정밀화하여 버킷 오버플로우 시에도 무결 슬라이싱 보장.
     • 문장 종결 부호(?, !, ...)에 따른 \sentence_types\ (1: 진술, 2: 의문, 3: 감탄, 4: 말줄임표) 및 어휘 시작 경계(\wordStarts\ -> \oundaryEvents) 매핑을 온전하게 전개하여 의문문·감탄문 억양 자연도 개선.
   • 검증:
     • 96개 음소(한국어 2문장, 227프레임) 합성 시 기존 2.422.75초에서 잘리던 오디오가 4.845.69초의 온전한 파형으로 전체 음소 시퀀스가 처음부터 끝까지 완전 재생됨을 검증.

9. Git Commit & Push 완료 및 Hugging Face Space 정식 배포:

   • 모델 리포지토리 동기화 (cidsound/scyllasband):
     • WebGPU 버퍼 락 순차화, Vocos LATENT_HOP 512 지속 시간 수정, 다중 경로 복원력(Multi-tier fallback) 및 Prosody 이벤트 강화 반영 완료.
     • 로컬 Git 변경 사항 원격 main 브랜치에 완전 커밋 및 푸시 완료 (2b1b4d3).
   • Hugging Face Space 배포 완료 (cidsound/scyllasband-webgpu):
     • Space 유형: Static Space (sdk: static — 무중단 24/7 가동, 콜드 스타트 0초, HF 글로벌 CDN 호스팅).
     • Space URL: https://huggingface.co/spaces/acidsound/scyllasband-webgpu
     • 직접 접속 URL (전체화면): https://acidsound-scyllasband-webgpu.static.hf.space
     • 230 MB 초경량 INT8 번들이 Space 내에 직접 배포되어 외부 의존성 없이 완전한 100% 자립 실행(Self-contained) 동작.
     • 배포 상태: Space stage: RUNNING, 모든 정적 파일, WASM 바이너리(pplication/wasm), 230 MB 가중치 및 ONNX 그래프 HTTP 200 서빙 검증 완료.