o xjjW@sdZddlmZddlZddlZddlZddlZddlmZddl m Z e dZ Gddde Zed d Zgd gd gd gdgddZgdZddddddZddddddZddZiddddd dd!dd"dd#dd$d$d%d$d&d$d'd$d(d$d)d)d*d)d+d+d,d+d-d+Zed.Z   dldmd6d7Zdnd:d;Zdod=d>Zed?d@ZdpdBdCZdqdrdGdHZdsdKdLZ     dtdudUdVZ!dvdwdYdZZ" [ \     dxdydjdkZ#dS)zuTTS backends for data pipeline. Two implementations: - ElevenLabs (paid, emotional prosody) — recommended - edge-tts (free fallback) Env vars: ELEVENLABS_API_KEY (required for elevenlabs) ELEVENLABS_VOICE_ID (optional, default to multilingual-friendly voice) ) annotationsN)Path)Listttsc@seZdZdZdS)QuotaExhausteduRaised when ElevenLabs returns 401/402/quota — non-retryable. synth_all catches this once and aborts the whole batch so the user can see a single clear message and re-run with a different key.N)__name__ __module__ __qualname____doc__r r L/dataset/kemix-engine/package/face/animasync-face-v3/scripts/compiler/tts.pyrsrELEVENLABS_MODEL eleven_v3)FCdKzv68Ofr4VUDcZXIyKsAmBSHXsuxsZ1lZZXlGoHB9Xhox1bqMl1TvkmelZubHeGTOAkECknc02Zmo)m3yAHyFEFKtbCIM5n7GFk9073AMdU5sAUtPMH1ilhWXqitL3DEOLD49pgNWRz2P4oCxSHhXan3ew4COv)TbMNBJ27fH2U0VgpSNko1W00IGEmNmwmsDeYy7agOEaq3WGNtNvFJ5co9mJExi3rF0t7dg7uN2M0WUhr)8tsLeAV5vPVuzCCvqbbUs0XGIcqmceN2l7kjsqoZiWLjl1zCuqXRkW6494ve8QclKarwLctvzN2plke3)XfNU2rGpBa01ckF309OY93nuHbke4dTER9x2pDwEYDDaC9XKjODs7hY78qEWBNr4zvrC1bGIdIstzjFQ)angersadnessjoyneutralsurprise)rrrrrrrrrrrrrrr c sdd|D}|s dSddlm}|}i|D]:}|dp d}t|d}||d7<|dp7gd }t|d krFtt|dnd }|d ||<qt|fd d| D} t| dkrn| dSt| fd ddS)uPick the most-frequent base emotion across a scenario's turns. `turns` is a list of dicts with 'emotion' (and optionally 'vad'). Returns the base name ('joy', 'sadness', etc.). Tie-break: highest summed |arousal| of the tied bases. Empty input → 'neutral'. cSs"g|] }|dp dr|qS)text)getstrip).0tr r r ps"z+dominant_base_for_turns..r&r)Counteremotionvad)r3r3r3csg|] \}}|kr|qSr r )r,bc) top_countr r r.~scs |dS)Nr3)r*)r5) arousal_sumr r s z)dominant_base_for_turns..)key) collectionsr/r*EMOTION_TO_BASElenabsfloatmaxvaluesitems) turnsrealr/countsr-emobaser2atiedr )r8r7r dominant_base_for_turnsgs$      rJr&r%laughter excitement agreement gratituder$cryingsulkapologystruggler#refusalr'flustershyELEVENLABS_VOICE_IDr0 str | Noneoverrideseed pool_overridereturnstrcCs~|r|StrtS|dkrt}n|rt|}n t|pdd}t|}|dur*|dStt|d d}||t |S)uPick voice ID from pool keyed by emotion → base emotion. seed: stable key (e.g. text or scenario_id+turn_idx) to deterministically pick a voice from the pool. Without seed, picks first voice in pool. Deterministic selection ensures same turn always gets same voice on re-run. pool_override: force voice selection from a specific base-emotion pool (e.g. 'neutral' for monologues — same voice across all emotions in turn). female_monologuer&Nrutf-8) SINGLE_VOICE_OVERRIDEFEMALE_MONOLOGUE_POOLVOICE_POOLS_BY_BASEr<r*inthashlibmd5encode hexdigestr=)r0rXrYrZpoolrGhr r r voice_id_for_emotions  rjr2 list | NonecCsv|dur dddddS|\}}}tt|}ttdtdd d |}ttd tddd |}d}|||ddS) uMap VAD → ElevenLabs voice_settings for expressive prosody. For v3 + audio tags: tags carry the emotion, voice_settings carry consistency. Stacking high `style` on top of expressive tags triple-amplifies expression and produces distortion (raspy / accent drift / foreigner-sounding) — that was the failure mode observed on 2026-05-07. We keep style modest and stability solid. Intensity is |A| magnitude — both high-positive (excited) and low-negative (sad) arousal scale prosody equally. Ng?g?g333333?T) stabilitysimilarity_booststyleuse_speaker_boostr3g?g?g?g?g?)r>r?r@min)r2VAD intensityrnrlrmr r r _build_voice_settingss   rur(cCs$|}|r|ddvr|dS|S)uFEnsure text ends with sentence-ending punctuation so v3 doesn't cut the tail. Note: `、` (U+3001) is a Japanese reading-comma, not a Korean sentence terminator — it was here by mistake. Korean uses ASCII `.?!` plus sometimes `…`. CJK terminators `。？！` are kept as a safety net for pasted text. u.?!。？！….)r+)r(r r r _ensure_trailing_punctsrxreu9[ㅠㅜ]{2,}|[ㅋㅎ]{2,}|[ㅏㅓㅗㅡㅣ]{2,}|[ㅡ]{2,}tuple[str, bool, bool]cCsnddl}t|d|}t|d|}|dd|}|dd|}|dd|}|d d|}|||fS) udv3 reads bare Korean jamo runs (ㅠㅠ, ㅋㅋ, ㅎㅎ) literally — comes out as "어어/유유/크크" instead of sobbing/laughing. Replace with v3 audio tags so they render as real reactions. ㅠㅠ / ㅜㅜ → [crying] (canonical v3 demo tag — fires reliably where [sniffles] was getting ignored mid-Korean-sentence) ㅋㅋ / ㅎㅎ → [laughs] (canonical v3 demo tag — fires reliably where [chuckles] was getting ignored) bare vowel runs (ㅏㅏ, ㅡㅡ…) → stripped (garbage) No leading period — that produced a hard sentence break before the reaction (audible "speak…long pause…laugh" instead of "speak laugh"). The canonical tags fire mid-clause without it. Returns (cleaned_text, had_sob, had_chuckle) — flags reserved but head-tag suppression is OFF by default now. Head tag carries the emotional prosody for the whole line; the body tag just stamps the reaction sound. Suppressing the head was making the voiced emotion drop out entirely. rNu [ㅠㅜ]{2,}u [ㅋㅎ]{2,}z [crying] z [laughs] u[ㅏㅓㅗㅡㅣ]{2,} z\s+)ryboolsearchsubr+)r(r0ryhad_sob had_chuckler r r _scrub_korean_emoticonss r min_charsrccCs"t||kr|ddS|S)uv3 fails / hallucinates on very short inputs (e.g. "응.", "네."). Pad with neutral filler so the model has enough context. We add an ellipsis-style trailer that v3 reads as a brief pause, not extra speech. rwz...)r=r+rstrip)r(rr r r _pad_short_textsr use_v3_tagsr|cCsddl}|dd|}t||\}}}t|}t|}|r#|r#|dkr%|Sdddddd d d d d dddddd}||dp>d}|rG|dvrGd}|rO|dvrOd}|r[d|d|S|S)uxPrepend emotion tag for v3 models (supports bracketed audio tags). Adds a leading period buffer before the tag — v3 sometimes eats the first token after an audio tag, which truncates the first word of real content. Also ensures trailing punctuation so the model doesn't cut the final word. Short texts are padded — v3 hallucinates on <12-char inputs. rNz \[[^\]]*\]r{r&z [happily]z[laughs]z [excited]z [sorrowful]z[crying]z[sighs]z[resigned tone]z [frustrated]z[angry]z[flatly]z [surprised]z [stammers]z [whispers])r%rKrLrMrNr$rOrPrQrRr#rSr'rTrUr))rKr%rL)r$rOrPrQz. )ryr~rrrxr*r+)r(r0rryrrtag_maptagr r r _format_text_with_emotion s:   rout_pathrvoice_idmodelapi_keylist voice_seed voice_poolc sddlmt||pj|dptptds#tddv} t ||| t |t } fdd} dd l } d } t| D] }z | d | Id HWd StyW}zt|d vpd vpd vpdvpdvpdvpdv}|rz#tdjddt|jdt|dddt|ddd ddWn tyYnwtdd d|dvpd v}d!vpd"vpd#v}tfd$d%d&D}tfd'd%d(D}|p|p|p|}|rR|| d)krR|rd*nd+}td,|d-|| dd.}|rd!n |r!dn|r&d/nd0}td1|d2d d3d4|d)d5| d6|d7d8 t |Id HWYd }~qMd }~wwd S)9u2Synthesize one clip with ElevenLabs, using emotion+VAD for prosody. If voice_id is None, picks voice automatically based on emotion → base mapping. voice_seed: stable key for deterministic voice pick from the pool. voice_pool: override base-emotion pool (e.g. 'neutral' for monologues). r) ElevenLabs)rXrYrZELEVENLABS_API_KEYz#ELEVENLABS_API_KEY env var requiredv3csddl}d}|jjdd}d}|d}|D]}||q!Wdn1s3wYz!|jddd d d t|d d ddddtg dd|jddWdSt |j fy}}zt d|d| WYd}~dSd}~ww)Nr)r mp3_44100_128)rr(model_id output_formatvoice_settingsz.raw.mp3wbffmpegz-yz -loglevelerrorz-iz-afz adelay=180|180,apad=pad_dur=0.18z-c:a libmp3lamez-b:a128kT)check) missing_okz[tts] ffmpeg pad failed (z); saving raw audio) subprocesstext_to_speechconvert with_suffixopenwriterunr\unlinkFileNotFoundErrorCalledProcessErrorLOGwarningreplace)rclientaudioraw_pathfchunke)rrformatted_textrrrrr r _sync^s>    z#synth_one_elevenlabs.._syncN401402403quota unauthorizedpayment_requiredbillingz/tmp/tts_quota_error.txtz=== full exception === z === type === z === status_code attr === status_codezn/az === body attr === body r^)encodingzElevenLabs auth/quota error: 409already_running429 rate_limittoo_many_requestsc3|]}|vVqdSNr r,r6)msgr r z'synth_one_elevenlabs..)z 500 z 502 z 503 z 504 c3rrr r)lowr r rr)timeout connectionzremote disconnectedr1g @g@gN@r4g?5xxnet[tts] z on voice u …, retry /z in z.1fs)elevenlabs.clientrrjstemDEFAULT_EL_MODELosgetenv RuntimeErrorlowerrruasyncioget_event_looprandomrangerun_in_executor Exceptionr\r write_texttypergetattrranyrpuniformrinfosleep)r(rrrrr0r2rrrlooprr max_retriesattemptris_quotais_409is_429is_5xxis_net retryablerGdelaykindr ) rrrrrrrrrr synth_one_elevenlabsAs   /   "8 rko-KR-SunHiNeuralvoicec sjtjdd|d|dt|tjjtjjd IdH}|IdH\}}|jdkr3td| dd dS) zFree fallback via edge-tts CLI.zedge-ttsz--voicez--textz --write-media)stdoutstderrNrzedge-tts failed: r) rcreate_subprocess_execr\rDEVNULLPIPE communicate returncoderdecode)r(rrproc_errr r r synth_one_edges  r elevenlabstexts List[str] out_paths List[Path]backend concurrencyemotionsvads List[list] voice_seeds voice_pools voice_ids List[bool]c st tks Jt|dgt dddddkr$tnt d fdd tjfd d tt DId Hr\td dt ddrntddd ddS)zParallel TTS for a batch. Returns success flag per item. emotions/vads: per-item lists aligned with texts. If provided, passed into the TTS call for emotional prosody. Frr))flagreasonrircc s4IdHdr WdIdHdS|}|r;|jdkr;d|<d7 WdIdHdSz`t}dkrdurN||d< durX ||d< durh |durh ||d< durx |durx ||d < dur |dur ||d < ||fi|IdHd|<Wn@ty}zdd<t|d <td |WYd}~n,d}~wty}zt d |d|WYd}~nd}~wwWdIdHdSWdIdHdSWdIdHdS1IdHswYdS)NriTr1rr0r2rrrru*[tts] QUOTA EXHAUSTED — aborting batch: z [tts] fail z: ) existsstatst_sizedictrr\rrrr)rpath per_kwargsr)abortedr r kwargsokr semaphoreskippedsynth_fnrr rrrr r onesR       "0zsynth_all..onecsg|]}|qSr r )r,r)r!r r r.szsynth_all..Nrrz turns skipped (already on disk)ru< [tts] ⚠ batch aborted on quota/auth error. reason: rruz partial results saved. Re-run with a different ELEVENLABS_API_KEY to resume — already-done turns will be skipped.)rrc)r=r Semaphorerrgatherrprint) rrr r r r rrrrr )rr r rrr!rrrr rr rrrr synth_alls   (&! r%)NNN) r0rWrXrWrYrWrZrWr[r\)r2rk)r(r\r[r\)r(r\r0rWr[rz)r)r(r\rrcr[r\)r(r\r0rWrr|r[r\)NNNNNNN)r(r\rrrr\rr\rr\r0r\r2rrr\rr\)r)r(r\rrrr\)rrNNNNN)rrrrr r\r rcr rr r rrrrrrr[r)$r __future__rrrdloggingrpathlibrtypingr getLoggerrrrrrrbraFEMALE_BY_BASE MALE_BY_BASErJr<r`rjrurx __import__compile_KOREAN_EMOTICON_RErrrrrr%r r r r s     %        4 u